| 1 |
|
| 2 | |
| 3 | * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
|
| 4 | *
|
| 5 | * Copyright (c) 2025 Meta Platforms, Inc. and affiliates.
|
| 6 | * Copyright (c) 2025 Tejun Heo <[email protected]>
|
| 7 |
|
| 8 | #ifndef _KERNEL_SCHED_EXT_INTERNAL_H
|
| 9 | #define _KERNEL_SCHED_EXT_INTERNAL_H
|
| 10 |
|
| 11 | #include "../sched.h"
|
| 12 | #include "types.h"
|
| 13 |
|
| 14 | #include <trace/events/sched_ext.h>
|
| 15 |
|
| 16 | |
| 17 | * scx_add_event - Increase an event counter for 'name' by 'cnt'
|
| 18 | * @sch: scx_sched to account events for
|
| 19 | * @name: an event name defined in struct scx_event_stats
|
| 20 | * @cnt: the number of the event occurred
|
| 21 | *
|
| 22 | * This can be used when preemption is not disabled.
|
| 23 |
|
| 24 | #define scx_add_event(sch, name, cnt) do { \
|
| 25 | this_cpu_add((sch)->pcpu->event_stats.name, (cnt)); \
|
| 26 | trace_sched_ext_event(#name, (cnt)); \
|
| 27 | } while(0)
|
| 28 |
|
| 29 | |
| 30 | * __scx_add_event - Increase an event counter for 'name' by 'cnt'
|
| 31 | * @sch: scx_sched to account events for
|
| 32 | * @name: an event name defined in struct scx_event_stats
|
| 33 | * @cnt: the number of the event occurred
|
| 34 | *
|
| 35 | * This should be used only when preemption is disabled.
|
| 36 |
|
| 37 | #define __scx_add_event(sch, name, cnt) do { \
|
| 38 | __this_cpu_add((sch)->pcpu->event_stats.name, (cnt)); \
|
| 39 | trace_sched_ext_event(#name, cnt); \
|
| 40 | } while(0)
|
| 41 |
|
| 42 | #define SCX_OP_IDX(op) (offsetof(struct sched_ext_ops, op) / sizeof(void (*)(void)))
|
| 43 | #define SCX_MOFF_IDX(moff) ((moff) / sizeof(void (*)(void)))
|
| 44 |
|
| 45 | enum scx_exit_kind {
|
| 46 | SCX_EXIT_NONE,
|
| 47 | SCX_EXIT_DONE,
|
| 48 |
|
| 49 | SCX_EXIT_UNREG = 64,
|
| 50 | SCX_EXIT_UNREG_BPF,
|
| 51 | SCX_EXIT_UNREG_KERN,
|
| 52 | SCX_EXIT_SYSRQ,
|
| 53 | SCX_EXIT_PARENT,
|
| 54 | SCX_EXIT_PARENT_KILL,
|
| 55 |
|
| 56 | SCX_EXIT_ERROR = 1024,
|
| 57 | SCX_EXIT_ERROR_BPF,
|
| 58 | SCX_EXIT_ERROR_STALL,
|
| 59 | SCX_EXIT_ERROR_REENQ,
|
| 60 | SCX_EXIT_ERROR_RESCUE,
|
| 61 | };
|
| 62 |
|
| 63 | |
| 64 | * An exit code can be specified when exiting with scx_bpf_exit() or scx_exit(),
|
| 65 | * corresponding to exit_kind UNREG_BPF and UNREG_KERN respectively. The codes
|
| 66 | * are 64bit of the format:
|
| 67 | *
|
| 68 | * Bits: [63 .. 48 47 .. 32 31 .. 0]
|
| 69 | * [ SYS ACT ] [ SYS RSN ] [ USR ]
|
| 70 | *
|
| 71 | * SYS ACT: System-defined exit actions
|
| 72 | * SYS RSN: System-defined exit reasons
|
| 73 | * USR : User-defined exit codes and reasons
|
| 74 | *
|
| 75 | * Using the above, users may communicate intention and context by ORing system
|
| 76 | * actions and/or system reasons with a user-defined exit code.
|
| 77 |
|
| 78 | enum scx_exit_code {
|
| 79 |
|
| 80 | SCX_ECODE_RSN_HOTPLUG = 1LLU << 32,
|
| 81 | SCX_ECODE_RSN_CGROUP_OFFLINE = 2LLU << 32,
|
| 82 |
|
| 83 |
|
| 84 | SCX_ECODE_ACT_RESTART = 1LLU << 48,
|
| 85 | };
|
| 86 |
|
| 87 | enum scx_exit_flags {
|
| 88 | |
| 89 | * ops.exit() may be called even if the loading failed before ops.init()
|
| 90 | * finishes successfully. This is because ops.exit() allows rich exit
|
| 91 | * info communication. The following flag indicates whether ops.init()
|
| 92 | * finished successfully.
|
| 93 |
|
| 94 | SCX_EFLAG_INITIALIZED = 1LLU << 0,
|
| 95 | };
|
| 96 |
|
| 97 | |
| 98 | * scx_exit_info is passed to ops.exit() to describe why the BPF scheduler is
|
| 99 | * being disabled.
|
| 100 |
|
| 101 | struct scx_exit_info {
|
| 102 |
|
| 103 | enum scx_exit_kind kind;
|
| 104 |
|
| 105 | |
| 106 | * CPU that initiated the exit, valid once @kind has been set.
|
| 107 | * Negative if the exit path didn't identify a CPU.
|
| 108 |
|
| 109 | s32 exit_cpu;
|
| 110 |
|
| 111 |
|
| 112 | s64 exit_code;
|
| 113 |
|
| 114 |
|
| 115 | u64 flags;
|
| 116 |
|
| 117 |
|
| 118 | const char *reason;
|
| 119 |
|
| 120 |
|
| 121 | unsigned long *bt;
|
| 122 | u32 bt_len;
|
| 123 |
|
| 124 |
|
| 125 | char *msg;
|
| 126 |
|
| 127 |
|
| 128 | char *dump;
|
| 129 | };
|
| 130 |
|
| 131 |
|
| 132 | enum scx_ops_flags {
|
| 133 | |
| 134 | * Keep built-in idle tracking even if ops.update_idle() is implemented.
|
| 135 |
|
| 136 | SCX_OPS_KEEP_BUILTIN_IDLE = 1LLU << 0,
|
| 137 |
|
| 138 | |
| 139 | * By default, if there are no other task to run on the CPU, ext core
|
| 140 | * keeps running the current task even after its slice expires. If this
|
| 141 | * flag is specified, such tasks are passed to ops.enqueue() with
|
| 142 | * %SCX_ENQ_LAST. See the comment above %SCX_ENQ_LAST for more info.
|
| 143 |
|
| 144 | SCX_OPS_ENQ_LAST = 1LLU << 1,
|
| 145 |
|
| 146 | |
| 147 | * An exiting task may schedule after PF_EXITING is set. In such cases,
|
| 148 | * bpf_task_from_pid() may not be able to find the task and if the BPF
|
| 149 | * scheduler depends on pid lookup for dispatching, the task will be
|
| 150 | * lost leading to various issues including RCU grace period stalls.
|
| 151 | *
|
| 152 | * To mask this problem, by default, unhashed tasks are automatically
|
| 153 | * dispatched to the local DSQ on enqueue. If the BPF scheduler doesn't
|
| 154 | * depend on pid lookups and wants to handle these tasks directly, the
|
| 155 | * following flag can be used. With %SCX_OPS_TID_TO_TASK,
|
| 156 | * scx_bpf_tid_to_task() can find exiting tasks reliably.
|
| 157 |
|
| 158 | SCX_OPS_ENQ_EXITING = 1LLU << 2,
|
| 159 |
|
| 160 | |
| 161 | * If set, only tasks with policy set to SCHED_EXT are attached to
|
| 162 | * sched_ext. If clear, SCHED_NORMAL tasks are also included.
|
| 163 |
|
| 164 | SCX_OPS_SWITCH_PARTIAL = 1LLU << 3,
|
| 165 |
|
| 166 | |
| 167 | * A migration disabled task can only execute on its current CPU. By
|
| 168 | * default, such tasks are automatically put on the CPU's local DSQ with
|
| 169 | * the default slice on enqueue. If this ops flag is set, they also go
|
| 170 | * through ops.enqueue().
|
| 171 | *
|
| 172 | * A migration disabled task never invokes ops.select_cpu() as it can
|
| 173 | * only select the current CPU. Also, p->cpus_ptr will only contain its
|
| 174 | * current CPU while p->nr_cpus_allowed keeps tracking p->user_cpus_ptr
|
| 175 | * and thus may disagree with cpumask_weight(p->cpus_ptr).
|
| 176 |
|
| 177 | SCX_OPS_ENQ_MIGRATION_DISABLED = 1LLU << 4,
|
| 178 |
|
| 179 | |
| 180 | * Queued wakeup (ttwu_queue) is a wakeup optimization that invokes
|
| 181 | * ops.enqueue() on the ops.select_cpu() selected or the wakee's
|
| 182 | * previous CPU via IPI (inter-processor interrupt) to reduce cacheline
|
| 183 | * transfers. When this optimization is enabled, ops.select_cpu() is
|
| 184 | * skipped in some cases (when racing against the wakee switching out).
|
| 185 | * As the BPF scheduler may depend on ops.select_cpu() being invoked
|
| 186 | * during wakeups, queued wakeup is disabled by default.
|
| 187 | *
|
| 188 | * If this ops flag is set, queued wakeup optimization is enabled and
|
| 189 | * the BPF scheduler must be able to handle ops.enqueue() invoked on the
|
| 190 | * wakee's CPU without preceding ops.select_cpu() even for tasks which
|
| 191 | * may be executed on multiple CPUs.
|
| 192 |
|
| 193 | SCX_OPS_ALLOW_QUEUED_WAKEUP = 1LLU << 5,
|
| 194 |
|
| 195 | |
| 196 | * If set, enable per-node idle cpumasks. If clear, use a single global
|
| 197 | * flat idle cpumask.
|
| 198 |
|
| 199 | SCX_OPS_BUILTIN_IDLE_PER_NODE = 1LLU << 6,
|
| 200 |
|
| 201 | |
| 202 | * If set, %SCX_ENQ_IMMED is assumed to be set on all local DSQ
|
| 203 | * enqueues.
|
| 204 |
|
| 205 | SCX_OPS_ALWAYS_ENQ_IMMED = 1LLU << 7,
|
| 206 |
|
| 207 | |
| 208 | * Maintain a mapping from p->scx.tid to task_struct so the BPF
|
| 209 | * scheduler can recover task pointers from stored tids via
|
| 210 | * scx_bpf_tid_to_task().
|
| 211 | *
|
| 212 | * Only the root scheduler turns this on. A sub-sched may set the flag
|
| 213 | * to declare a dependency on the lookup; if the root scheduler hasn't
|
| 214 | * enabled it, attaching the sub-sched is rejected.
|
| 215 |
|
| 216 | SCX_OPS_TID_TO_TASK = 1LLU << 8,
|
| 217 |
|
| 218 | SCX_OPS_ALL_FLAGS = SCX_OPS_KEEP_BUILTIN_IDLE |
|
| 219 | SCX_OPS_ENQ_LAST |
|
| 220 | SCX_OPS_ENQ_EXITING |
|
| 221 | SCX_OPS_ENQ_MIGRATION_DISABLED |
|
| 222 | SCX_OPS_ALLOW_QUEUED_WAKEUP |
|
| 223 | SCX_OPS_SWITCH_PARTIAL |
|
| 224 | SCX_OPS_BUILTIN_IDLE_PER_NODE |
|
| 225 | SCX_OPS_ALWAYS_ENQ_IMMED |
|
| 226 | SCX_OPS_TID_TO_TASK,
|
| 227 |
|
| 228 |
|
| 229 | __SCX_OPS_INTERNAL_MASK = 0xffLLU << 56,
|
| 230 |
|
| 231 | SCX_OPS_HAS_CPU_PREEMPT = 1LLU << 56,
|
| 232 | };
|
| 233 |
|
| 234 |
|
| 235 | struct scx_init_task_args {
|
| 236 | |
| 237 | * Set if ops.init_task() is being invoked on the fork path, as opposed
|
| 238 | * to the scheduler transition path.
|
| 239 |
|
| 240 | bool fork;
|
| 241 | #ifdef CONFIG_EXT_GROUP_SCHED
|
| 242 |
|
| 243 | struct cgroup *cgroup;
|
| 244 | #endif
|
| 245 | };
|
| 246 |
|
| 247 |
|
| 248 | struct scx_exit_task_args {
|
| 249 |
|
| 250 | bool cancelled;
|
| 251 | };
|
| 252 |
|
| 253 | |
| 254 | * struct scx_enable_args - Argument container for cid-form ops.enable()
|
| 255 | * @cmask_arena_addr: BPF arena address of the cmask of cids the task may run on
|
| 256 | *
|
| 257 | * @cmask_arena_addr is the task's affinity as it enters the scheduler.
|
| 258 | * set_cmask() delivers the same mask right after enable(), before set_weight()
|
| 259 | * and the first enqueue, then every affinity change afterwards, and is never
|
| 260 | * called before enable(). A scheduler may therefore track affinity in
|
| 261 | * set_cmask() alone.
|
| 262 | *
|
| 263 | * The kernel builds the mask in the scheduler arena from its own geometry, so
|
| 264 | * the header is valid regardless of what the scheduler last wrote there. The
|
| 265 | * memory is per-cpu scratch reused once the callback returns: copy the bits
|
| 266 | * out, don't keep the address. The set_cmask() argument follows the same rules.
|
| 267 | *
|
| 268 | * The address is a plain value rather than a typed pointer because BTF can't
|
| 269 | * mark a struct member as an arena pointer yet and a pointer member would reach
|
| 270 | * the program typed as a kernel pointer. Cast it to struct scx_cmask __arena *
|
| 271 | * before use. Once arena members can be typed, a typed alias will join this
|
| 272 | * field in an anonymous union at the same offset.
|
| 273 |
|
| 274 | struct scx_enable_args {
|
| 275 | u64 cmask_arena_addr;
|
| 276 | };
|
| 277 |
|
| 278 |
|
| 279 | struct scx_cgroup_init_args {
|
| 280 |
|
| 281 | u32 weight;
|
| 282 |
|
| 283 |
|
| 284 | u64 bw_period_us;
|
| 285 | u64 bw_quota_us;
|
| 286 | u64 bw_burst_us;
|
| 287 |
|
| 288 |
|
| 289 | bool sched_idle;
|
| 290 | };
|
| 291 |
|
| 292 | enum scx_cpu_preempt_reason {
|
| 293 |
|
| 294 | SCX_CPU_PREEMPT_RT,
|
| 295 |
|
| 296 | SCX_CPU_PREEMPT_DL,
|
| 297 |
|
| 298 | SCX_CPU_PREEMPT_STOP,
|
| 299 |
|
| 300 | SCX_CPU_PREEMPT_UNKNOWN,
|
| 301 | };
|
| 302 |
|
| 303 | |
| 304 | * Argument container for ops.cpu_acquire(). Currently empty, but may be
|
| 305 | * expanded in the future.
|
| 306 |
|
| 307 | struct scx_cpu_acquire_args {};
|
| 308 |
|
| 309 |
|
| 310 | struct scx_cpu_release_args {
|
| 311 |
|
| 312 | enum scx_cpu_preempt_reason reason;
|
| 313 |
|
| 314 |
|
| 315 | struct task_struct *task;
|
| 316 | };
|
| 317 |
|
| 318 |
|
| 319 | struct scx_dump_ctx {
|
| 320 | enum scx_exit_kind kind;
|
| 321 | s64 exit_code;
|
| 322 | const char *reason;
|
| 323 | u64 at_ns;
|
| 324 | u64 at_jiffies;
|
| 325 | };
|
| 326 |
|
| 327 |
|
| 328 | struct scx_sub_attach_args {
|
| 329 | struct sched_ext_ops *ops;
|
| 330 | char *cgroup_path;
|
| 331 | };
|
| 332 |
|
| 333 |
|
| 334 | struct scx_sub_detach_args {
|
| 335 | struct sched_ext_ops *ops;
|
| 336 | char *cgroup_path;
|
| 337 | };
|
| 338 |
|
| 339 | |
| 340 | * struct sched_ext_ops - Operation table for BPF scheduler implementation
|
| 341 | *
|
| 342 | * A BPF scheduler can implement an arbitrary scheduling policy by
|
| 343 | * implementing and loading operations in this table. Note that a userland
|
| 344 | * scheduling policy can also be implemented using the BPF scheduler
|
| 345 | * as a shim layer.
|
| 346 |
|
| 347 | struct sched_ext_ops {
|
| 348 | |
| 349 | * @select_cpu: Pick the target CPU for a task which is being woken up
|
| 350 | * @p: task being woken up
|
| 351 | * @prev_cpu: the cpu @p was on before sleeping
|
| 352 | * @wake_flags: SCX_WAKE_*
|
| 353 | *
|
| 354 | * Decision made here isn't final. @p may be moved to any CPU while it
|
| 355 | * is getting dispatched for execution later. However, as @p is not on
|
| 356 | * the rq at this point, getting the eventual execution CPU right here
|
| 357 | * saves a small bit of overhead down the line.
|
| 358 | *
|
| 359 | * If an idle CPU is returned, the CPU is kicked and will try to
|
| 360 | * dispatch. While an explicit custom mechanism can be added,
|
| 361 | * select_cpu() serves as the default way to wake up idle CPUs.
|
| 362 | *
|
| 363 | * @p may be inserted into a DSQ directly by calling
|
| 364 | * scx_bpf_dsq_insert(). If so, the ops.enqueue() will be skipped.
|
| 365 | * Directly inserting into %SCX_DSQ_LOCAL will put @p in the local DSQ
|
| 366 | * of the CPU returned by this operation.
|
| 367 | *
|
| 368 | * Note that select_cpu() is never called for tasks that can only run
|
| 369 | * on a single CPU or tasks with migration disabled, as they don't have
|
| 370 | * the option to select a different CPU. See select_task_rq() for
|
| 371 | * details.
|
| 372 |
|
| 373 | s32 (*select_cpu)(struct task_struct *p, s32 prev_cpu, u64 wake_flags);
|
| 374 |
|
| 375 | |
| 376 | * @enqueue: Enqueue a task on the BPF scheduler
|
| 377 | * @p: task being enqueued
|
| 378 | * @enq_flags: %SCX_ENQ_*
|
| 379 | *
|
| 380 | * @p is ready to run. Insert directly into a DSQ by calling
|
| 381 | * scx_bpf_dsq_insert() or enqueue on the BPF scheduler. If not directly
|
| 382 | * inserted, the bpf scheduler owns @p and if it fails to dispatch @p,
|
| 383 | * the task will stall.
|
| 384 | *
|
| 385 | * If @p was inserted into a DSQ from ops.select_cpu(), this callback is
|
| 386 | * skipped.
|
| 387 |
|
| 388 | void (*enqueue)(struct task_struct *p, u64 enq_flags);
|
| 389 |
|
| 390 | |
| 391 | * @dequeue: Remove a task from the BPF scheduler
|
| 392 | * @p: task being dequeued
|
| 393 | * @deq_flags: %SCX_DEQ_*
|
| 394 | *
|
| 395 | * Remove @p from the BPF scheduler. This is usually called to isolate
|
| 396 | * the task while updating its scheduling properties (e.g. priority).
|
| 397 | *
|
| 398 | * The ext core keeps track of whether the BPF side owns a given task or
|
| 399 | * not and can gracefully ignore spurious dispatches from BPF side,
|
| 400 | * which makes it safe to not implement this method. However, depending
|
| 401 | * on the scheduling logic, this can lead to confusing behaviors - e.g.
|
| 402 | * scheduling position not being updated across a priority change.
|
| 403 |
|
| 404 | void (*dequeue)(struct task_struct *p, u64 deq_flags);
|
| 405 |
|
| 406 | |
| 407 | * @dispatch: Dispatch tasks from the BPF scheduler and/or user DSQs
|
| 408 | * @cpu: CPU to dispatch tasks for
|
| 409 | * @prev: previous task being switched out
|
| 410 | *
|
| 411 | * Called when a CPU's local dsq is empty. The operation should dispatch
|
| 412 | * one or more tasks from the BPF scheduler into the DSQs using
|
| 413 | * scx_bpf_dsq_insert() and/or move from user DSQs into the local DSQ
|
| 414 | * using scx_bpf_dsq_move_to_local().
|
| 415 | *
|
| 416 | * The maximum number of times scx_bpf_dsq_insert() can be called
|
| 417 | * without an intervening scx_bpf_dsq_move_to_local() is specified by
|
| 418 | * ops.dispatch_max_batch. See the comments on top of the two functions
|
| 419 | * for more details.
|
| 420 | *
|
| 421 | * When not %NULL, @prev is an SCX task with its slice depleted. If
|
| 422 | * @prev is still runnable as indicated by set %SCX_TASK_QUEUED in
|
| 423 | * @prev->scx.flags, it is not enqueued yet and will be enqueued after
|
| 424 | * ops.dispatch() returns. To keep executing @prev, return without
|
| 425 | * dispatching or moving any tasks. Also see %SCX_OPS_ENQ_LAST.
|
| 426 |
|
| 427 | void (*dispatch)(s32 cpu, struct task_struct *prev);
|
| 428 |
|
| 429 | |
| 430 | * @tick: Periodic tick
|
| 431 | * @p: task running currently
|
| 432 | *
|
| 433 | * This operation is called every 1/HZ seconds on CPUs which are
|
| 434 | * executing an SCX task. Setting a slice of 0 for @p with
|
| 435 | * scx_bpf_task_set_slice() will trigger an immediate dispatch cycle on
|
| 436 | * the CPU.
|
| 437 |
|
| 438 | void (*tick)(struct task_struct *p);
|
| 439 |
|
| 440 | |
| 441 | * @runnable: A task is becoming runnable on its associated CPU
|
| 442 | * @p: task becoming runnable
|
| 443 | * @enq_flags: %SCX_ENQ_*
|
| 444 | *
|
| 445 | * This and the following three functions can be used to track a task's
|
| 446 | * execution state transitions. A task becomes ->runnable() on a CPU,
|
| 447 | * and then goes through one or more ->running() and ->stopping() pairs
|
| 448 | * as it runs on the CPU, and eventually becomes ->quiescent() when it's
|
| 449 | * done running on the CPU.
|
| 450 | *
|
| 451 | * @p is becoming runnable on the CPU because it's
|
| 452 | *
|
| 453 | * - waking up (%SCX_ENQ_WAKEUP)
|
| 454 | * - being moved from another CPU
|
| 455 | * - being restored after temporarily taken off the queue for an
|
| 456 | * attribute change.
|
| 457 | *
|
| 458 | * This and ->enqueue() are related but not coupled. This operation
|
| 459 | * notifies @p's state transition and may not be followed by ->enqueue()
|
| 460 | * e.g. when @p is being dispatched to a remote CPU, or when @p is
|
| 461 | * being enqueued on a CPU experiencing a hotplug event. Likewise, a
|
| 462 | * task may be ->enqueue()'d without being preceded by this operation
|
| 463 | * e.g. after exhausting its slice.
|
| 464 |
|
| 465 | void (*runnable)(struct task_struct *p, u64 enq_flags);
|
| 466 |
|
| 467 | |
| 468 | * @running: A task is starting to run on its associated CPU
|
| 469 | * @p: task starting to run
|
| 470 | *
|
| 471 | * Note that this callback may be called from a CPU other than the
|
| 472 | * one the task is going to run on. This can happen when a task
|
| 473 | * property is changed (i.e., affinity), since set_next_task_scx(),
|
| 474 | * which triggers this callback, may run on a CPU different from
|
| 475 | * the task's assigned CPU.
|
| 476 | *
|
| 477 | * Therefore, always use scx_bpf_task_cpu(@p) to determine the
|
| 478 | * target CPU the task is going to use.
|
| 479 | *
|
| 480 | * See ->runnable() for explanation on the task state notifiers.
|
| 481 |
|
| 482 | void (*running)(struct task_struct *p);
|
| 483 |
|
| 484 | |
| 485 | * @stopping: A task is stopping execution
|
| 486 | * @p: task stopping to run
|
| 487 | * @runnable: is task @p still runnable?
|
| 488 | *
|
| 489 | * Note that this callback may be called from a CPU other than the
|
| 490 | * one the task was running on. This can happen when a task
|
| 491 | * property is changed (i.e., affinity), since dequeue_task_scx(),
|
| 492 | * which triggers this callback, may run on a CPU different from
|
| 493 | * the task's assigned CPU.
|
| 494 | *
|
| 495 | * Therefore, always use scx_bpf_task_cpu(@p) to retrieve the CPU
|
| 496 | * the task was running on.
|
| 497 | *
|
| 498 | * See ->runnable() for explanation on the task state notifiers. If
|
| 499 | * !@runnable, ->quiescent() will be invoked after this operation
|
| 500 | * returns.
|
| 501 |
|
| 502 | void (*stopping)(struct task_struct *p, bool runnable);
|
| 503 |
|
| 504 | |
| 505 | * @quiescent: A task is becoming not runnable on its associated CPU
|
| 506 | * @p: task becoming not runnable
|
| 507 | * @deq_flags: %SCX_DEQ_*
|
| 508 | *
|
| 509 | * See ->runnable() for explanation on the task state notifiers.
|
| 510 | *
|
| 511 | * @p is becoming quiescent on the CPU because it's
|
| 512 | *
|
| 513 | * - sleeping (%SCX_DEQ_SLEEP)
|
| 514 | * - being moved to another CPU
|
| 515 | * - being temporarily taken off the queue for an attribute change
|
| 516 | * (%SCX_DEQ_SCHED_CHANGE)
|
| 517 | *
|
| 518 | * This and ->dequeue() are related but not coupled. This operation
|
| 519 | * notifies @p's state transition and may not be preceded by ->dequeue()
|
| 520 | * e.g. when @p is being dispatched to a remote CPU.
|
| 521 |
|
| 522 | void (*quiescent)(struct task_struct *p, u64 deq_flags);
|
| 523 |
|
| 524 | |
| 525 | * @yield: Yield CPU
|
| 526 | * @from: yielding task
|
| 527 | * @to: optional yield target task
|
| 528 | *
|
| 529 | * If @to is NULL, @from is yielding the CPU to other runnable tasks.
|
| 530 | * The BPF scheduler should ensure that other available tasks are
|
| 531 | * dispatched before the yielding task. Return value is ignored in this
|
| 532 | * case.
|
| 533 | *
|
| 534 | * If @to is not-NULL, @from wants to yield the CPU to @to. If the bpf
|
| 535 | * scheduler can implement the request, return %true; otherwise, %false.
|
| 536 |
|
| 537 | bool (*yield)(struct task_struct *from, struct task_struct *to);
|
| 538 |
|
| 539 | |
| 540 | * @core_sched_before: Task ordering for core-sched
|
| 541 | * @a: task A
|
| 542 | * @b: task B
|
| 543 | *
|
| 544 | * Used by core-sched to determine the ordering between two tasks. See
|
| 545 | * Documentation/admin-guide/hw-vuln/core-scheduling.rst for details on
|
| 546 | * core-sched.
|
| 547 | *
|
| 548 | * Both @a and @b are runnable and may or may not currently be queued on
|
| 549 | * the BPF scheduler. Should return %true if @a should run before @b.
|
| 550 | * %false if there's no required ordering or @b should run before @a.
|
| 551 | *
|
| 552 | * In a scheduler hierarchy, a pair spanning two schedulers is ordered
|
| 553 | * by the nearest common ancestor implementing this op, so the op may be
|
| 554 | * called on tasks that the scheduler delegated to its sub-schedulers
|
| 555 | * and is not scheduling anymore. See scx_prio_less().
|
| 556 | *
|
| 557 | * If not specified, the default is ordering them according to when they
|
| 558 | * became runnable.
|
| 559 |
|
| 560 | bool (*core_sched_before)(struct task_struct *a, struct task_struct *b);
|
| 561 |
|
| 562 | |
| 563 | * @set_weight: Set task weight
|
| 564 | * @p: task to set weight for
|
| 565 | * @weight: new weight [1..10000]
|
| 566 | *
|
| 567 | * Update @p's weight to @weight.
|
| 568 |
|
| 569 | void (*set_weight)(struct task_struct *p, u32 weight);
|
| 570 |
|
| 571 | |
| 572 | * @set_cpumask: Set CPU affinity
|
| 573 | * @p: task to set CPU affinity for
|
| 574 | * @cpumask: cpumask of cpus that @p can run on
|
| 575 | *
|
| 576 | * Update @p's CPU affinity to @cpumask.
|
| 577 |
|
| 578 | void (*set_cpumask)(struct task_struct *p,
|
| 579 | const struct cpumask *cpumask);
|
| 580 |
|
| 581 | |
| 582 | * @update_idle: Update the idle state of a CPU
|
| 583 | * @cpu: CPU to update the idle state for
|
| 584 | * @idle: whether entering or exiting the idle state
|
| 585 | *
|
| 586 | * This operation is called when @rq's CPU goes or leaves the idle
|
| 587 | * state. By default, implementing this operation disables the built-in
|
| 588 | * idle CPU tracking and the following helpers become unavailable:
|
| 589 | *
|
| 590 | * - scx_bpf_select_cpu_dfl()
|
| 591 | * - scx_bpf_select_cpu_and()
|
| 592 | * - scx_bpf_test_and_clear_cpu_idle()
|
| 593 | * - scx_bpf_pick_idle_cpu()
|
| 594 | *
|
| 595 | * The user also must implement ops.select_cpu() as the default
|
| 596 | * implementation relies on scx_bpf_select_cpu_dfl().
|
| 597 | *
|
| 598 | * Specify the %SCX_OPS_KEEP_BUILTIN_IDLE flag to keep the built-in idle
|
| 599 | * tracking.
|
| 600 | *
|
| 601 | * Only actual transitions are reported. A CPU that is claimed with an
|
| 602 | * idle pick and kicked but dispatches no task returns to idle without a
|
| 603 | * transition. A scheduler tracking idle CPUs itself must restore the
|
| 604 | * idle state from ops.dispatch() when it returns without the next task
|
| 605 | * to run.
|
| 606 |
|
| 607 | void (*update_idle)(s32 cpu, bool idle);
|
| 608 |
|
| 609 | |
| 610 | * @init_task: Initialize a task to run in a BPF scheduler
|
| 611 | * @p: task to initialize for BPF scheduling
|
| 612 | * @args: init arguments, see the struct definition
|
| 613 | *
|
| 614 | * Either we're loading a BPF scheduler or a new task is being forked.
|
| 615 | * Initialize @p for BPF scheduling. This operation may block and can
|
| 616 | * be used for allocations, and is called exactly once for a task.
|
| 617 | *
|
| 618 | * Return 0 for success, -errno for failure. An error return while
|
| 619 | * loading will abort loading of the BPF scheduler. During a fork, it
|
| 620 | * will abort that specific fork.
|
| 621 |
|
| 622 | s32 (*init_task)(struct task_struct *p, struct scx_init_task_args *args);
|
| 623 |
|
| 624 | |
| 625 | * @exit_task: Exit a previously-running task from the system
|
| 626 | * @p: task to exit
|
| 627 | * @args: exit arguments, see the struct definition
|
| 628 | *
|
| 629 | * @p is exiting or the BPF scheduler is being unloaded. Perform any
|
| 630 | * necessary cleanup for @p.
|
| 631 |
|
| 632 | void (*exit_task)(struct task_struct *p, struct scx_exit_task_args *args);
|
| 633 |
|
| 634 | |
| 635 | * @enable: Enable BPF scheduling for a task
|
| 636 | * @p: task to enable BPF scheduling for
|
| 637 | *
|
| 638 | * Enable @p for BPF scheduling. enable() is called on @p any time it
|
| 639 | * enters SCX, and is always paired with a matching disable().
|
| 640 |
|
| 641 | void (*enable)(struct task_struct *p);
|
| 642 |
|
| 643 | |
| 644 | * @disable: Disable BPF scheduling for a task
|
| 645 | * @p: task to disable BPF scheduling for
|
| 646 | *
|
| 647 | * @p is exiting, leaving SCX or the BPF scheduler is being unloaded.
|
| 648 | * Disable BPF scheduling for @p. A disable() call is always matched
|
| 649 | * with a prior enable() call.
|
| 650 |
|
| 651 | void (*disable)(struct task_struct *p);
|
| 652 |
|
| 653 | |
| 654 | * @dump: Dump BPF scheduler state on error
|
| 655 | * @ctx: debug dump context
|
| 656 | *
|
| 657 | * Use scx_bpf_dump() to generate BPF scheduler specific debug dump.
|
| 658 |
|
| 659 | void (*dump)(struct scx_dump_ctx *ctx);
|
| 660 |
|
| 661 | |
| 662 | * @dump_cpu: Dump BPF scheduler state for a CPU on error
|
| 663 | * @ctx: debug dump context
|
| 664 | * @cpu: CPU to generate debug dump for
|
| 665 | * @idle: @cpu is currently idle without any runnable tasks
|
| 666 | *
|
| 667 | * Use scx_bpf_dump() to generate BPF scheduler specific debug dump for
|
| 668 | * @cpu. If @idle is %true and this operation doesn't produce any
|
| 669 | * output, @cpu is skipped for dump.
|
| 670 |
|
| 671 | void (*dump_cpu)(struct scx_dump_ctx *ctx, s32 cpu, bool idle);
|
| 672 |
|
| 673 | |
| 674 | * @dump_task: Dump BPF scheduler state for a runnable task on error
|
| 675 | * @ctx: debug dump context
|
| 676 | * @p: runnable task to generate debug dump for
|
| 677 | *
|
| 678 | * Use scx_bpf_dump() to generate BPF scheduler specific debug dump for
|
| 679 | * @p.
|
| 680 |
|
| 681 | void (*dump_task)(struct scx_dump_ctx *ctx, struct task_struct *p);
|
| 682 |
|
| 683 | #ifdef CONFIG_EXT_GROUP_SCHED
|
| 684 | |
| 685 | * @cgroup_init: Initialize a cgroup
|
| 686 | * @cgrp: cgroup being initialized
|
| 687 | * @args: init arguments, see the struct definition
|
| 688 | *
|
| 689 | * Initialize @cgrp for sched_ext, delivered to @cgrp's sched either
|
| 690 | * when the BPF scheduler is being loaded or when @cgrp is created. This
|
| 691 | * operation may block.
|
| 692 | *
|
| 693 | * Cgroup handovers also generate these ops: an enabling sub-scheduler
|
| 694 | * receives ops.cgroup_init() for every cgroup in its subtree while the
|
| 695 | * previous sched receives ops.cgroup_exit(), and disabling reverses the
|
| 696 | * two.
|
| 697 | *
|
| 698 | * When the BPF scheduler is being loaded or cgroups are being handed
|
| 699 | * over, @cgrp may already have been removed by userspace: a removed
|
| 700 | * cgroup stays schedulable until its dying tasks finish their final
|
| 701 | * context switches.
|
| 702 | *
|
| 703 | * Return 0 for success, -errno for failure. An error return while
|
| 704 | * loading will abort loading of the BPF scheduler. During cgroup
|
| 705 | * creation, it will abort the specific cgroup creation.
|
| 706 |
|
| 707 | s32 (*cgroup_init)(struct cgroup *cgrp,
|
| 708 | struct scx_cgroup_init_args *args);
|
| 709 |
|
| 710 | |
| 711 | * @cgroup_exit: Exit a cgroup
|
| 712 | * @cgrp: cgroup being exited
|
| 713 | *
|
| 714 | * Exit @cgrp for sched_ext, delivered to the sched whose
|
| 715 | * ops.cgroup_init() it pairs with, either when the BPF scheduler is
|
| 716 | * being unloaded or when @cgrp is destroyed. This operation may block.
|
| 717 | *
|
| 718 | * For a destroyed @cgrp, delivery follows the last scheduling event on
|
| 719 | * it: a removed cgroup stays schedulable until its dying tasks finish
|
| 720 | * their final context switches.
|
| 721 |
|
| 722 | void (*cgroup_exit)(struct cgroup *cgrp);
|
| 723 |
|
| 724 | |
| 725 | * @cgroup_prep_move: Prepare a task to be moved to a different cgroup
|
| 726 | * @p: task being moved
|
| 727 | * @from: cgroup @p is being moved from
|
| 728 | * @to: cgroup @p is being moved to
|
| 729 | *
|
| 730 | * Prepare @p for move from cgroup @from to @to. This operation may
|
| 731 | * block and can be used for allocations.
|
| 732 | *
|
| 733 | * The cgroup_move ops are delivered to @p's sched, and only for moves
|
| 734 | * that don't re-home @p. A re-homing move is reported through
|
| 735 | * ops.exit_task() and ops.init_task() instead. @from and @to can
|
| 736 | * reference cgroups the sched never received ops.cgroup_init() for, as
|
| 737 | * the cpu controller can be coarser than the sub-scheduler topology.
|
| 738 | *
|
| 739 | * Return 0 for success, -errno for failure. An error return aborts the
|
| 740 | * migration.
|
| 741 |
|
| 742 | s32 (*cgroup_prep_move)(struct task_struct *p,
|
| 743 | struct cgroup *from, struct cgroup *to);
|
| 744 |
|
| 745 | |
| 746 | * @cgroup_move: Commit cgroup move
|
| 747 | * @p: task being moved
|
| 748 | * @from: cgroup @p is being moved from
|
| 749 | * @to: cgroup @p is being moved to
|
| 750 | *
|
| 751 | * Commit the move. @p is dequeued during this operation.
|
| 752 |
|
| 753 | void (*cgroup_move)(struct task_struct *p,
|
| 754 | struct cgroup *from, struct cgroup *to);
|
| 755 |
|
| 756 | |
| 757 | * @cgroup_cancel_move: Cancel cgroup move
|
| 758 | * @p: task whose cgroup move is being canceled
|
| 759 | * @from: cgroup @p was being moved from
|
| 760 | * @to: cgroup @p was being moved to
|
| 761 | *
|
| 762 | * @p was cgroup_prep_move()'d but failed before reaching cgroup_move().
|
| 763 | * Undo the preparation.
|
| 764 |
|
| 765 | void (*cgroup_cancel_move)(struct task_struct *p,
|
| 766 | struct cgroup *from, struct cgroup *to);
|
| 767 |
|
| 768 | |
| 769 | * @cgroup_set_weight: A cgroup's weight is being changed
|
| 770 | * @cgrp: cgroup whose weight is being updated
|
| 771 | * @weight: new weight [1..10000]
|
| 772 | *
|
| 773 | * Update @cgrp's weight to @weight.
|
| 774 | *
|
| 775 | * Knobs of a cgroup belong to the parent, so the set_* ops are
|
| 776 | * delivered to @cgrp's parent's sched. That sched may never have seen
|
| 777 | * ops.cgroup_init() for @cgrp - at a sub-scheduler attach point, the
|
| 778 | * parent sched tracks @cgrp through ops.sub_attach() instead.
|
| 779 |
|
| 780 | void (*cgroup_set_weight)(struct cgroup *cgrp, u32 weight);
|
| 781 |
|
| 782 | |
| 783 | * @cgroup_set_bandwidth: A cgroup's bandwidth is being changed
|
| 784 | * @cgrp: cgroup whose bandwidth is being updated
|
| 785 | * @period_us: bandwidth control period
|
| 786 | * @quota_us: bandwidth control quota
|
| 787 | * @burst_us: bandwidth control burst
|
| 788 | *
|
| 789 | * Update @cgrp's bandwidth control parameters. This is from the cpu.max
|
| 790 | * cgroup interface. This operation may block.
|
| 791 | *
|
| 792 | * @quota_us / @period_us determines the CPU bandwidth @cgrp is entitled
|
| 793 | * to. For example, if @period_us is 1_000_000 and @quota_us is
|
| 794 | * 2_500_000. @cgrp is entitled to 2.5 CPUs. @burst_us can be
|
| 795 | * interpreted in the same fashion and specifies how much @cgrp can
|
| 796 | * burst temporarily. The specific control mechanism and thus the
|
| 797 | * interpretation of @period_us and burstiness is up to the BPF
|
| 798 | * scheduler.
|
| 799 | *
|
| 800 | * Delivery follows the same rule as cgroup_set_weight().
|
| 801 |
|
| 802 | void (*cgroup_set_bandwidth)(struct cgroup *cgrp,
|
| 803 | u64 period_us, u64 quota_us, u64 burst_us);
|
| 804 |
|
| 805 | |
| 806 | * @cgroup_set_idle: A cgroup's idle state is being changed
|
| 807 | * @cgrp: cgroup whose idle state is being updated
|
| 808 | * @idle: whether the cgroup is entering or exiting idle state
|
| 809 | *
|
| 810 | * Update @cgrp's idle state to @idle. This callback is invoked when
|
| 811 | * a cgroup transitions between idle and non-idle states, allowing the
|
| 812 | * BPF scheduler to adjust its behavior accordingly.
|
| 813 | *
|
| 814 | * Delivery follows the same rule as cgroup_set_weight().
|
| 815 |
|
| 816 | void (*cgroup_set_idle)(struct cgroup *cgrp, bool idle);
|
| 817 |
|
| 818 | #endif
|
| 819 |
|
| 820 | |
| 821 | * @sub_attach: Attach a sub-scheduler
|
| 822 | * @args: argument container, see the struct definition
|
| 823 | *
|
| 824 | * Return 0 to accept the sub-scheduler. -errno to reject.
|
| 825 |
|
| 826 | s32 (*sub_attach)(struct scx_sub_attach_args *args);
|
| 827 |
|
| 828 | |
| 829 | * @sub_detach: Detach a sub-scheduler
|
| 830 | * @args: argument container, see the struct definition
|
| 831 |
|
| 832 | void (*sub_detach)(struct scx_sub_detach_args *args);
|
| 833 |
|
| 834 | |
| 835 | * @sub_caps_updated: Caps on this sub-sched's shard changed
|
| 836 | * @cmask: cids whose caps changed (cmask->base identifies the shard)
|
| 837 | * @caps: SCX_CAP_* that changed
|
| 838 | *
|
| 839 | * Invoked after grant or revoke modifies caps on a shard. There can be
|
| 840 | * only one in-flight invocation per shard. @cmask and @caps coalesce
|
| 841 | * all changes since the last delivery. Direction (set vs cleared) isn't
|
| 842 | * encoded. Query current state with scx_bpf_sub_caps().
|
| 843 | *
|
| 844 | * Delivered asynchronously after the change is recorded, and may run
|
| 845 | * before it takes effect on any given cpu. Use it to track which caps
|
| 846 | * the sub-sched holds and propagate to its own children, not to decide
|
| 847 | * if a task can run on a cpu now. sub_ecaps_updated() reports that per
|
| 848 | * cpu, once it is in effect.
|
| 849 | *
|
| 850 | * May call scx_bpf_sub_grant() / scx_bpf_sub_revoke() on children.
|
| 851 |
|
| 852 | void (*sub_caps_updated)(const struct scx_cmask *cmask, u64 caps);
|
| 853 |
|
| 854 | |
| 855 | * @sub_ecaps_updated: This sub-sched's effective caps on a cid changed
|
| 856 | * @cid: the cid whose effective caps changed
|
| 857 | * @before: effective caps as of the last delivery
|
| 858 | * @after: effective caps now
|
| 859 | *
|
| 860 | * Invoked when this sub-sched's effective caps on @cid change, once the
|
| 861 | * change is in effect on the cpu. Runs in dispatch context with rq lock
|
| 862 | * held, and can perform all operations allowed in ops.dispatch()
|
| 863 | * including inserting/moving tasks.
|
| 864 |
|
| 865 | void (*sub_ecaps_updated)(s32 cid, u64 before, u64 after);
|
| 866 |
|
| 867 | |
| 868 | * All online ops must come before ops.cpu_online().
|
| 869 |
|
| 870 |
|
| 871 | |
| 872 | * @cpu_online: A CPU became online
|
| 873 | * @cpu: CPU which just came up
|
| 874 | *
|
| 875 | * @cpu just came online. @cpu will not call ops.enqueue() or
|
| 876 | * ops.dispatch(), nor run tasks associated with other CPUs beforehand.
|
| 877 |
|
| 878 | void (*cpu_online)(s32 cpu);
|
| 879 |
|
| 880 | |
| 881 | * @cpu_offline: A CPU is going offline
|
| 882 | * @cpu: CPU which is going offline
|
| 883 | *
|
| 884 | * @cpu is going offline. @cpu will not call ops.enqueue() or
|
| 885 | * ops.dispatch(), nor run tasks associated with other CPUs afterwards.
|
| 886 |
|
| 887 | void (*cpu_offline)(s32 cpu);
|
| 888 |
|
| 889 | |
| 890 | * All CPU hotplug ops must come before ops.init_cids().
|
| 891 |
|
| 892 |
|
| 893 | |
| 894 | * @init_cids: Finalize the cid layout (cid-form only)
|
| 895 | *
|
| 896 | * Runs after the default cid layout is built, before caps and shards
|
| 897 | * are finalized. A cid-form scheduler may call scx_bpf_cid_override()
|
| 898 | * here for a custom layout. Ignored for cpu-form schedulers.
|
| 899 |
|
| 900 | s32 (*init_cids)(void);
|
| 901 |
|
| 902 | |
| 903 | * @init: Initialize the BPF scheduler
|
| 904 |
|
| 905 | s32 (*init)(void);
|
| 906 |
|
| 907 | |
| 908 | * @exit: Clean up after the BPF scheduler
|
| 909 | * @info: Exit info
|
| 910 | *
|
| 911 | * ops.exit() is also called on ops.init() failure, which is a bit
|
| 912 | * unusual. This is to allow rich reporting through @info on how
|
| 913 | * ops.init() failed.
|
| 914 |
|
| 915 | void (*exit)(struct scx_exit_info *info);
|
| 916 |
|
| 917 | |
| 918 | * Data fields must comes after all ops fields.
|
| 919 |
|
| 920 |
|
| 921 | |
| 922 | * @dispatch_max_batch: Max nr of tasks that dispatch() can dispatch
|
| 923 |
|
| 924 | u32 dispatch_max_batch;
|
| 925 |
|
| 926 | |
| 927 | * @flags: %SCX_OPS_* flags
|
| 928 |
|
| 929 | u64 flags;
|
| 930 |
|
| 931 | |
| 932 | * @timeout_ms: The maximum amount of time, in milliseconds, that a
|
| 933 | * runnable task should be able to wait before being scheduled. The
|
| 934 | * maximum timeout may not exceed the default timeout of 30 seconds.
|
| 935 | *
|
| 936 | * Defaults to the maximum allowed timeout value of 30 seconds.
|
| 937 |
|
| 938 | u32 timeout_ms;
|
| 939 |
|
| 940 | |
| 941 | * @exit_dump_len: scx_exit_info.dump buffer length. If 0, the default
|
| 942 | * value of 32768 is used.
|
| 943 |
|
| 944 | u32 exit_dump_len;
|
| 945 |
|
| 946 | |
| 947 | * @hotplug_seq: A sequence number that may be set by the scheduler to
|
| 948 | * detect when a hotplug event has occurred during the loading process.
|
| 949 | * If 0, no detection occurs. Otherwise, the scheduler will fail to
|
| 950 | * load if the sequence number does not match @scx_hotplug_seq on the
|
| 951 | * enable path.
|
| 952 |
|
| 953 | u64 hotplug_seq;
|
| 954 |
|
| 955 | |
| 956 | * @cid_shard_size: Target number of CIDs per shard
|
| 957 | *
|
| 958 | * Shards are contiguous CID ranges used as operation and locking
|
| 959 | * domains for sub-scheduling. Each LLC is divided into ceil(nr_cpus /
|
| 960 | * @cid_shard_size) shards, then cores are distributed across them
|
| 961 | * evenly. If one core has more logical CPUs than @cid_shard_size, its
|
| 962 | * shard will become larger than @cid_shard_size. Values above
|
| 963 | * SCX_CID_SHARD_MAX_CPUS are capped. 0 means use the default (24).
|
| 964 |
|
| 965 | u32 cid_shard_size;
|
| 966 |
|
| 967 | |
| 968 | * @rescue_bandwidth_ppt: Rescue execution bandwidth in parts per thousand
|
| 969 | *
|
| 970 | * The fraction of each CPU's time that may be consumed running tasks
|
| 971 | * from its rescue DSQ. A higher bandwidth admits and escalates rescues
|
| 972 | * faster, see @rescue_quantum_us.
|
| 973 | *
|
| 974 | * Only the root scheduler's value is used. 0 means the default of 20
|
| 975 | * (2%). May not exceed 250 (25%). %SCX_RESCUE_DISABLE disables rescue -
|
| 976 | * %SCX_ENQ_RESCUE inserts are then rejected like any other insert
|
| 977 | * lacking the caps.
|
| 978 |
|
| 979 | u32 rescue_bandwidth_ppt;
|
| 980 |
|
| 981 | |
| 982 | * @rescue_quantum_us: Rescue execution quantum in microseconds
|
| 983 | *
|
| 984 | * How much CPU time each rescue gets. Rescues run one at a time per CPU
|
| 985 | * and admissions are paced to keep rescue execution within
|
| 986 | * @rescue_bandwidth_ppt - with the defaults, one 5ms rescue every
|
| 987 | * 250ms. A crowded queue round-robins on the quantum divided across the
|
| 988 | * waiters, floored at 1ms. A stuck rescue eventually escalates to
|
| 989 | * forced execution. A larger quantum interrupts the CPU less often but
|
| 990 | * for longer and spaces rescues further apart.
|
| 991 | *
|
| 992 | * Only the root scheduler's value is used. 0 means the default (5000).
|
| 993 | * Non-zero values must be within [1000, 100000]. Values too short for
|
| 994 | * the kernel to meter are lifted silently.
|
| 995 |
|
| 996 | u32 rescue_quantum_us;
|
| 997 |
|
| 998 | |
| 999 | * @sub_cgroup_id: When >1, attach the scheduler as a sub-scheduler
|
| 1000 | * on the specified cgroup.
|
| 1001 |
|
| 1002 | u64 sub_cgroup_id;
|
| 1003 |
|
| 1004 | |
| 1005 | * @name: BPF scheduler's name
|
| 1006 | *
|
| 1007 | * Must be a non-zero valid BPF object name including only isalnum(),
|
| 1008 | * '_' and '.' chars. Exposed via the ops file in the scheduler's sysfs
|
| 1009 | * directory, /sys/kernel/sched_ext/root/ops for the root scheduler,
|
| 1010 | * while the BPF scheduler is enabled.
|
| 1011 |
|
| 1012 | char name[SCX_OPS_NAME_LEN];
|
| 1013 |
|
| 1014 |
|
| 1015 | void __rcu *priv;
|
| 1016 |
|
| 1017 | |
| 1018 | * Deprecated callbacks. Kept at the end of the struct so the cid-form
|
| 1019 | * struct (sched_ext_ops_cid) can omit them without affecting the
|
| 1020 | * shared field offsets. Use SCX_ENQ_IMMED instead. Sitting past
|
| 1021 | * SCX_OPI_END means has_op doesn't cover them, so SCX_HAS_OP() cannot
|
| 1022 | * be used; callers must test sch->ops.cpu_acquire / cpu_release
|
| 1023 | * directly.
|
| 1024 |
|
| 1025 |
|
| 1026 | |
| 1027 | * @cpu_acquire: A CPU is becoming available to the BPF scheduler
|
| 1028 | * @cpu: The CPU being acquired by the BPF scheduler.
|
| 1029 | * @args: Acquire arguments, see the struct definition.
|
| 1030 | *
|
| 1031 | * A CPU that was previously released from the BPF scheduler is now once
|
| 1032 | * again under its control. Deprecated; use SCX_ENQ_IMMED instead.
|
| 1033 |
|
| 1034 | void (*cpu_acquire)(s32 cpu, struct scx_cpu_acquire_args *args);
|
| 1035 |
|
| 1036 | |
| 1037 | * @cpu_release: A CPU is taken away from the BPF scheduler
|
| 1038 | * @cpu: The CPU being released by the BPF scheduler.
|
| 1039 | * @args: Release arguments, see the struct definition.
|
| 1040 | *
|
| 1041 | * The specified CPU is no longer under the control of the BPF
|
| 1042 | * scheduler. This could be because it was preempted by a higher
|
| 1043 | * priority sched_class, though there may be other reasons as well. The
|
| 1044 | * caller should consult @args->reason to determine the cause.
|
| 1045 | * Deprecated; use SCX_ENQ_IMMED instead.
|
| 1046 |
|
| 1047 | void (*cpu_release)(s32 cpu, struct scx_cpu_release_args *args);
|
| 1048 | };
|
| 1049 |
|
| 1050 | |
| 1051 | * struct sched_ext_ops_cid - cid-form alternative to struct sched_ext_ops
|
| 1052 | *
|
| 1053 | * Mirrors struct sched_ext_ops with cpu/cpumask substituted with cid/cmask
|
| 1054 | * where applicable. Layout up to and including @priv matches sched_ext_ops
|
| 1055 | * byte-for-byte (verified by BUILD_BUG_ON checks at scx_init() time) so
|
| 1056 | * shared field offsets work for both struct types in bpf_scx_init_member()
|
| 1057 | * and bpf_scx_check_member(). The deprecated cpu_acquire/cpu_release
|
| 1058 | * callbacks at the tail of sched_ext_ops are omitted here entirely.
|
| 1059 | *
|
| 1060 | * Differences from sched_ext_ops:
|
| 1061 | * - select_cpu -> select_cid (returns cid)
|
| 1062 | * - dispatch -> dispatch (cpu arg is now cid)
|
| 1063 | * - update_idle -> update_idle (cpu arg is now cid)
|
| 1064 | * - set_cpumask -> set_cmask (cmask instead of cpumask)
|
| 1065 | * - enable -> enable (takes struct scx_enable_args)
|
| 1066 | * - cpu_online -> cid_online
|
| 1067 | * - cpu_offline -> cid_offline
|
| 1068 | * - dump_cpu -> dump_cid
|
| 1069 | * - cgroup_* -> cpuctl_* (they track the cgroup cpu controller)
|
| 1070 | * - cpu_acquire/cpu_release -> not present (deprecated in sched_ext_ops)
|
| 1071 | *
|
| 1072 | * BPF schedulers using this type cannot call cpu-form scx_bpf_* kfuncs;
|
| 1073 | * use the cid-form variants instead. Enforced at BPF verifier time via
|
| 1074 | * scx_kfunc_context_filter() branching on prog->aux->st_ops.
|
| 1075 | *
|
| 1076 | * See sched_ext_ops for callback documentation.
|
| 1077 |
|
| 1078 | struct sched_ext_ops_cid {
|
| 1079 | s32 (*select_cid)(struct task_struct *p, s32 prev_cid, u64 wake_flags);
|
| 1080 | void (*enqueue)(struct task_struct *p, u64 enq_flags);
|
| 1081 | void (*dequeue)(struct task_struct *p, u64 deq_flags);
|
| 1082 | void (*dispatch)(s32 cid, struct task_struct *prev);
|
| 1083 | void (*tick)(struct task_struct *p);
|
| 1084 | void (*runnable)(struct task_struct *p, u64 enq_flags);
|
| 1085 | void (*running)(struct task_struct *p);
|
| 1086 | void (*stopping)(struct task_struct *p, bool runnable);
|
| 1087 | void (*quiescent)(struct task_struct *p, u64 deq_flags);
|
| 1088 | bool (*yield)(struct task_struct *from, struct task_struct *to);
|
| 1089 | bool (*core_sched_before)(struct task_struct *a,
|
| 1090 | struct task_struct *b);
|
| 1091 | void (*set_weight)(struct task_struct *p, u32 weight);
|
| 1092 | void (*set_cmask)(struct task_struct *p,
|
| 1093 | const struct scx_cmask *cmask__arena);
|
| 1094 | void (*update_idle)(s32 cid, bool idle);
|
| 1095 | s32 (*init_task)(struct task_struct *p,
|
| 1096 | struct scx_init_task_args *args);
|
| 1097 | void (*exit_task)(struct task_struct *p,
|
| 1098 | struct scx_exit_task_args *args);
|
| 1099 | void (*enable)(struct task_struct *p, struct scx_enable_args *args);
|
| 1100 | void (*disable)(struct task_struct *p);
|
| 1101 | void (*dump)(struct scx_dump_ctx *ctx);
|
| 1102 | void (*dump_cid)(struct scx_dump_ctx *ctx, s32 cid, bool idle);
|
| 1103 | void (*dump_task)(struct scx_dump_ctx *ctx, struct task_struct *p);
|
| 1104 | #ifdef CONFIG_EXT_GROUP_SCHED
|
| 1105 | s32 (*cpuctl_init)(struct cgroup *cgrp, struct scx_cgroup_init_args *args);
|
| 1106 | void (*cpuctl_exit)(struct cgroup *cgrp);
|
| 1107 | s32 (*cpuctl_prep_move)(struct task_struct *p, struct cgroup *from,
|
| 1108 | struct cgroup *to);
|
| 1109 | void (*cpuctl_move)(struct task_struct *p, struct cgroup *from, struct cgroup *to);
|
| 1110 | void (*cpuctl_cancel_move)(struct task_struct *p, struct cgroup *from,
|
| 1111 | struct cgroup *to);
|
| 1112 | void (*cpuctl_set_weight)(struct cgroup *cgrp, u32 weight);
|
| 1113 | void (*cpuctl_set_bandwidth)(struct cgroup *cgrp, u64 period_us, u64 quota_us,
|
| 1114 | u64 burst_us);
|
| 1115 | void (*cpuctl_set_idle)(struct cgroup *cgrp, bool idle);
|
| 1116 | #endif
|
| 1117 | s32 (*sub_attach)(struct scx_sub_attach_args *args);
|
| 1118 | void (*sub_detach)(struct scx_sub_detach_args *args);
|
| 1119 | void (*sub_caps_updated)(const struct scx_cmask *cmask__arena, u64 caps);
|
| 1120 | void (*sub_ecaps_updated)(s32 cid, u64 before, u64 after);
|
| 1121 | void (*cid_online)(s32 cid);
|
| 1122 | void (*cid_offline)(s32 cid);
|
| 1123 | s32 (*init_cids)(void);
|
| 1124 | s32 (*init)(void);
|
| 1125 | void (*exit)(struct scx_exit_info *info);
|
| 1126 |
|
| 1127 |
|
| 1128 | u32 dispatch_max_batch;
|
| 1129 | u64 flags;
|
| 1130 | u32 timeout_ms;
|
| 1131 | u32 exit_dump_len;
|
| 1132 | u64 hotplug_seq;
|
| 1133 | u32 cid_shard_size;
|
| 1134 | u32 rescue_bandwidth_ppt;
|
| 1135 | u32 rescue_quantum_us;
|
| 1136 | u64 sub_cgroup_id;
|
| 1137 | char name[SCX_OPS_NAME_LEN];
|
| 1138 |
|
| 1139 |
|
| 1140 | void __rcu *priv;
|
| 1141 |
|
| 1142 |
|
| 1143 | char __end[0];
|
| 1144 | };
|
| 1145 |
|
| 1146 | enum scx_opi {
|
| 1147 | SCX_OPI_BEGIN = 0,
|
| 1148 | SCX_OPI_NORMAL_BEGIN = 0,
|
| 1149 | SCX_OPI_NORMAL_END = SCX_OP_IDX(cpu_online),
|
| 1150 | SCX_OPI_CPU_HOTPLUG_BEGIN = SCX_OP_IDX(cpu_online),
|
| 1151 | SCX_OPI_CPU_HOTPLUG_END = SCX_OP_IDX(init_cids),
|
| 1152 | SCX_OPI_END = SCX_OP_IDX(init_cids),
|
| 1153 | };
|
| 1154 |
|
| 1155 | |
| 1156 | * Collection of event counters. Event types are placed in descending order.
|
| 1157 |
|
| 1158 | struct scx_event_stats {
|
| 1159 | |
| 1160 | * If ops.select_cpu() returns a CPU which can't be used by the task,
|
| 1161 | * the core scheduler code silently picks a fallback CPU.
|
| 1162 |
|
| 1163 | s64 SCX_EV_SELECT_CPU_FALLBACK;
|
| 1164 |
|
| 1165 | |
| 1166 | * When dispatching to a local DSQ, the CPU may have gone offline in
|
| 1167 | * the meantime. In this case, the task is bounced to the global DSQ.
|
| 1168 |
|
| 1169 | s64 SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE;
|
| 1170 |
|
| 1171 | |
| 1172 | * If SCX_OPS_ENQ_LAST is not set, the number of times that a task
|
| 1173 | * continued to run because there were no other tasks on the CPU.
|
| 1174 |
|
| 1175 | s64 SCX_EV_DISPATCH_KEEP_LAST;
|
| 1176 |
|
| 1177 | |
| 1178 | * If SCX_OPS_ENQ_EXITING is not set, the number of times that a task
|
| 1179 | * is dispatched to a local DSQ when exiting.
|
| 1180 |
|
| 1181 | s64 SCX_EV_ENQ_SKIP_EXITING;
|
| 1182 |
|
| 1183 | |
| 1184 | * If SCX_OPS_ENQ_MIGRATION_DISABLED is not set, the number of times a
|
| 1185 | * migration disabled task skips ops.enqueue() and is dispatched to its
|
| 1186 | * local DSQ.
|
| 1187 |
|
| 1188 | s64 SCX_EV_ENQ_SKIP_MIGRATION_DISABLED;
|
| 1189 |
|
| 1190 | |
| 1191 | * The number of times a task, enqueued on a local DSQ with
|
| 1192 | * SCX_ENQ_IMMED, was re-enqueued because the CPU was not available for
|
| 1193 | * immediate execution.
|
| 1194 |
|
| 1195 | s64 SCX_EV_REENQ_IMMED;
|
| 1196 |
|
| 1197 | |
| 1198 | * The number of times a reenqueue (%SCX_ENQ_REENQ) led to another
|
| 1199 | * reenqueue without the task running in between. This count climbing
|
| 1200 | * rapidly indicates that the BPF scheduler keeps re-deciding placements
|
| 1201 | * it can't honor. A single task reenqueued more than
|
| 1202 | * %SCX_REENQ_MAX_REPEAT times gets its owning scheduler ejected.
|
| 1203 |
|
| 1204 | s64 SCX_EV_REENQ_REPEAT;
|
| 1205 |
|
| 1206 | |
| 1207 | * Total number of times a task's time slice was refilled with the
|
| 1208 | * default value (SCX_SLICE_DFL).
|
| 1209 |
|
| 1210 | s64 SCX_EV_REFILL_SLICE_DFL;
|
| 1211 |
|
| 1212 | |
| 1213 | * The number of times an out-of-band slice request exceeded the maximum
|
| 1214 | * representable value and was clamped.
|
| 1215 |
|
| 1216 | s64 SCX_EV_SLICE_CLAMPED;
|
| 1217 |
|
| 1218 | |
| 1219 | * The number of times a slice extension was denied because the
|
| 1220 | * scheduler lacked baseline cpu access on the task's cpu.
|
| 1221 |
|
| 1222 | s64 SCX_EV_SLICE_DENIED;
|
| 1223 |
|
| 1224 | |
| 1225 | * The total duration of bypass modes in nanoseconds.
|
| 1226 |
|
| 1227 | s64 SCX_EV_BYPASS_DURATION;
|
| 1228 |
|
| 1229 | |
| 1230 | * The number of tasks dispatched in the bypassing mode.
|
| 1231 |
|
| 1232 | s64 SCX_EV_BYPASS_DISPATCH;
|
| 1233 |
|
| 1234 | |
| 1235 | * The number of times the bypassing mode has been activated.
|
| 1236 |
|
| 1237 | s64 SCX_EV_BYPASS_ACTIVATE;
|
| 1238 |
|
| 1239 | |
| 1240 | * The number of times the scheduler attempted to insert a task that it
|
| 1241 | * doesn't own into a DSQ. Such attempts are ignored.
|
| 1242 | *
|
| 1243 | * As BPF schedulers are allowed to ignore dequeues, it's difficult to
|
| 1244 | * tell whether such an attempt is from a scheduler malfunction or an
|
| 1245 | * ignored dequeue around sub-sched enabling. If this count keeps going
|
| 1246 | * up regardless of sub-sched enabling, it likely indicates a bug in the
|
| 1247 | * scheduler.
|
| 1248 |
|
| 1249 | s64 SCX_EV_INSERT_NOT_OWNED;
|
| 1250 |
|
| 1251 | |
| 1252 | * The number of times tasks from bypassing descendants are scheduled
|
| 1253 | * from sub_bypass_dsq's.
|
| 1254 |
|
| 1255 | s64 SCX_EV_SUB_BYPASS_DISPATCH;
|
| 1256 |
|
| 1257 | |
| 1258 | * The number of times a migration-disabled task lacking the cap for its
|
| 1259 | * cid was allowed onto the local DSQ. It must run on its pinned CPU, so
|
| 1260 | * it can't be rejected. The violation is counted here.
|
| 1261 |
|
| 1262 | s64 SCX_EV_SUB_FORCED_ADMIT;
|
| 1263 |
|
| 1264 | |
| 1265 | * The number of times a preempting kick was refused because the
|
| 1266 | * sub-sched lacked SCX_CAP_PREEMPT for a task outside its subtree. The
|
| 1267 | * kick degrades to a plain reschedule.
|
| 1268 |
|
| 1269 | s64 SCX_EV_SUB_PREEMPT_DENIED;
|
| 1270 |
|
| 1271 | |
| 1272 | * The number of times a kick was skipped because the sub-sched lacked
|
| 1273 | * baseline access on the target cid. The preempt-part degradation of a
|
| 1274 | * delivered kick is counted in SCX_EV_SUB_PREEMPT_DENIED instead.
|
| 1275 |
|
| 1276 | s64 SCX_EV_SUB_KICK_DENIED;
|
| 1277 |
|
| 1278 | |
| 1279 | * The number of times a local DSQ reenq was dropped because the
|
| 1280 | * sub-sched lacked baseline access on the target cid.
|
| 1281 |
|
| 1282 | s64 SCX_EV_SUB_REENQ_DENIED;
|
| 1283 |
|
| 1284 | |
| 1285 | * The number of times scx_bpf_cidperf_set() was denied because the
|
| 1286 | * sub-sched lacked SCX_CAP_PERF on the target cid.
|
| 1287 |
|
| 1288 | s64 SCX_EV_SUB_CIDPERF_DENIED;
|
| 1289 |
|
| 1290 | |
| 1291 | * The number of times an insert carrying %SCX_ENQ_RESCUE lacked the
|
| 1292 | * caps for its cid and the task entered the rescue path.
|
| 1293 |
|
| 1294 | s64 SCX_EV_SUB_RESCUE;
|
| 1295 | };
|
| 1296 |
|
| 1297 | #define SCX_EVENTS_LIST(SCX_EVENT) \
|
| 1298 | SCX_EVENT(SCX_EV_SELECT_CPU_FALLBACK); \
|
| 1299 | SCX_EVENT(SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE); \
|
| 1300 | SCX_EVENT(SCX_EV_DISPATCH_KEEP_LAST); \
|
| 1301 | SCX_EVENT(SCX_EV_ENQ_SKIP_EXITING); \
|
| 1302 | SCX_EVENT(SCX_EV_ENQ_SKIP_MIGRATION_DISABLED); \
|
| 1303 | SCX_EVENT(SCX_EV_REENQ_IMMED); \
|
| 1304 | SCX_EVENT(SCX_EV_REENQ_REPEAT); \
|
| 1305 | SCX_EVENT(SCX_EV_REFILL_SLICE_DFL); \
|
| 1306 | SCX_EVENT(SCX_EV_SLICE_CLAMPED); \
|
| 1307 | SCX_EVENT(SCX_EV_SLICE_DENIED); \
|
| 1308 | SCX_EVENT(SCX_EV_BYPASS_DURATION); \
|
| 1309 | SCX_EVENT(SCX_EV_BYPASS_DISPATCH); \
|
| 1310 | SCX_EVENT(SCX_EV_BYPASS_ACTIVATE); \
|
| 1311 | SCX_EVENT(SCX_EV_INSERT_NOT_OWNED); \
|
| 1312 | SCX_EVENT(SCX_EV_SUB_BYPASS_DISPATCH); \
|
| 1313 | SCX_EVENT(SCX_EV_SUB_FORCED_ADMIT); \
|
| 1314 | SCX_EVENT(SCX_EV_SUB_PREEMPT_DENIED); \
|
| 1315 | SCX_EVENT(SCX_EV_SUB_KICK_DENIED); \
|
| 1316 | SCX_EVENT(SCX_EV_SUB_REENQ_DENIED); \
|
| 1317 | SCX_EVENT(SCX_EV_SUB_CIDPERF_DENIED); \
|
| 1318 | SCX_EVENT(SCX_EV_SUB_RESCUE)
|
| 1319 |
|
| 1320 | struct scx_sched;
|
| 1321 |
|
| 1322 | enum scx_sched_pcpu_flags {
|
| 1323 | SCX_SCHED_PCPU_BYPASSING = 1LLU << 0,
|
| 1324 | };
|
| 1325 |
|
| 1326 |
|
| 1327 | struct scx_dsp_buf_ent {
|
| 1328 | struct task_struct *task;
|
| 1329 | unsigned long qseq;
|
| 1330 | u64 dsq_id;
|
| 1331 | u64 slice;
|
| 1332 | u64 vtime;
|
| 1333 | u64 enq_flags;
|
| 1334 | };
|
| 1335 |
|
| 1336 | struct scx_dsp_ctx {
|
| 1337 | struct rq *rq;
|
| 1338 | u32 cursor;
|
| 1339 | u32 nr_tasks;
|
| 1340 | struct scx_dsp_buf_ent buf[];
|
| 1341 | };
|
| 1342 |
|
| 1343 | struct scx_deferred_reenq_local {
|
| 1344 | struct list_head node;
|
| 1345 | u64 flags;
|
| 1346 | };
|
| 1347 |
|
| 1348 | struct scx_sched_pcpu {
|
| 1349 | struct scx_sched *sch;
|
| 1350 | u64 flags;
|
| 1351 |
|
| 1352 | |
| 1353 | * Kick state owned by this cpu for this sched. scx_kick_cpu() records
|
| 1354 | * targets here and links @to_kick_node onto the cpu's
|
| 1355 | * rq->scx.sched_pcpus_to_kick. The cpu's single kick irq_work walks
|
| 1356 | * that list and kicks each sched's targets on its behalf. Per-sched so
|
| 1357 | * a kick stays attributed to its scheduler.
|
| 1358 |
|
| 1359 | cpumask_var_t cpus_to_kick;
|
| 1360 | cpumask_var_t cpus_to_kick_if_idle;
|
| 1361 | cpumask_var_t cpus_to_preempt;
|
| 1362 | cpumask_var_t cpus_to_wait;
|
| 1363 | struct list_head to_kick_node;
|
| 1364 |
|
| 1365 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 1366 | |
| 1367 | * pshard->caps[cap_bit] is the set of cids the sched holds that one
|
| 1368 | * cap on. ecaps is its transpose: the set of SCX_CAP_* bits the sched
|
| 1369 | * effectively holds on this cpu, with implied caps folded in, so that
|
| 1370 | * the hot-path check is a single read.
|
| 1371 | *
|
| 1372 | * While pshard->caps[] under pshard->lock is the target configuration,
|
| 1373 | * ecaps is the effective copy owned by the cpu. It is written under the
|
| 1374 | * rq lock while processing rq->ecaps_to_sync. Can also be read with
|
| 1375 | * READ_ONCE() outside rq lock.
|
| 1376 | *
|
| 1377 | * See queue_sync_ecaps() and scx_process_sync_ecaps().
|
| 1378 |
|
| 1379 | u64 ecaps;
|
| 1380 | struct llist_node ecaps_to_sync_node;
|
| 1381 |
|
| 1382 | bool idle_renotify;
|
| 1383 |
|
| 1384 | u64 reported_ecaps;
|
| 1385 |
|
| 1386 | |
| 1387 | * Decaying rescue runtime consumed on this cpu, see
|
| 1388 | * scx_rescue_decay_avg(). Overload on this cpu ejects the sub with the
|
| 1389 | * largest value. Accessed only under this cpu's rq lock.
|
| 1390 |
|
| 1391 | u64 rescue_avg;
|
| 1392 | u64 rescue_avg_at;
|
| 1393 | #endif
|
| 1394 |
|
| 1395 | |
| 1396 | * The event counters are in a per-CPU variable to minimize the
|
| 1397 | * accounting overhead. A system-wide view on the event counter is
|
| 1398 | * constructed when requested by scx_bpf_events().
|
| 1399 |
|
| 1400 | struct scx_event_stats event_stats;
|
| 1401 |
|
| 1402 | struct scx_deferred_reenq_local deferred_reenq_local;
|
| 1403 | struct scx_dispatch_q bypass_dsq;
|
| 1404 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 1405 | u32 bypass_host_seq;
|
| 1406 | #endif
|
| 1407 |
|
| 1408 |
|
| 1409 | struct scx_dsp_ctx dsp_ctx;
|
| 1410 | };
|
| 1411 |
|
| 1412 | struct scx_sched_pnode {
|
| 1413 | struct scx_dispatch_q global_dsq;
|
| 1414 | };
|
| 1415 |
|
| 1416 | |
| 1417 | * Sub-sched capability delegation.
|
| 1418 | *
|
| 1419 | * Caps are per-cid permissions parents delegate to direct children via
|
| 1420 | * scx_bpf_sub_grant() / scx_bpf_sub_revoke(). A child's cap set is always a
|
| 1421 | * subset of its parent's. A sub-sched checks its caps locally, and cross-sched
|
| 1422 | * communication is needed only when the delegation set itself changes.
|
| 1423 | *
|
| 1424 | * Caps are used to implement sub-sched scheduling on the enqueue path. Picking
|
| 1425 | * a cid for a task at a leaf depends on which cids the leaf is allowed to use.
|
| 1426 | * Resolving that programmatically on every enqueue would mean a cross-sched
|
| 1427 | * round-trip call chain, possibly retrying if the request can't be granted
|
| 1428 | * as-is.
|
| 1429 | *
|
| 1430 | * The dispatch path is different - it runs as top-down recursion via
|
| 1431 | * scx_bpf_sub_dispatch(): a sched's dispatch op invokes a child's dispatch op
|
| 1432 | * on the local rq, and the subtree dispatches in a single pass.
|
| 1433 | *
|
| 1434 | * Locking is per shard. cid space is split into shards, and each sub-sched has
|
| 1435 | * its own pshard->lock for each shard. Operations are broken up on shard
|
| 1436 | * boundaries. Different shards never contend. Shards are expected to be
|
| 1437 | * topology-aligned and likely to serve as the locality unit when cids are
|
| 1438 | * allocated to schedulers, so per-shard lock granularity scales naturally with
|
| 1439 | * the allocation pattern.
|
| 1440 | *
|
| 1441 | * ENQ_IMMED insert an IMMED task onto the cid's local DSQ
|
| 1442 | * - kick the cid's cpu (except SCX_KICK_PREEMPT)
|
| 1443 | *
|
| 1444 | * ENQ insert any task onto the cid's local DSQ (implies ENQ_IMMED)
|
| 1445 | *
|
| 1446 | * PREEMPT preempt any task running on the cid regardless of the owning
|
| 1447 | * sched (implies ENQ). Preempting a task in the sched's own subtree
|
| 1448 | * doesn't require any cap.
|
| 1449 | * - SCX_ENQ_PREEMPT inserts
|
| 1450 | * - SCX_KICK_PREEMPT kicks
|
| 1451 | *
|
| 1452 | * PERF control the cid's cpu power/perf management state, currently the
|
| 1453 | * cpufreq target set through scx_bpf_cidperf_set(). Hardware
|
| 1454 | * control is a separate axis from queue access: PERF neither
|
| 1455 | * implies nor is implied by the caps above.
|
| 1456 | *
|
| 1457 | * Implied caps apply to the holder's own use of a cid, not to delegation.
|
| 1458 | * scx_bpf_sub_grant() delegates literally-held caps, so a cap held only through
|
| 1459 | * implication is usable but cannot be re-delegated to a child. When granting a
|
| 1460 | * cap, it usually makes sense to delegate its implied caps explicitly alongside
|
| 1461 | * it.
|
| 1462 |
|
| 1463 | enum scx_cap_flags {
|
| 1464 | __SCX_CAP_ENQ_IMMED = 0,
|
| 1465 | __SCX_CAP_ENQ = 1,
|
| 1466 | __SCX_CAP_PREEMPT = 2,
|
| 1467 | __SCX_CAP_PERF = 3,
|
| 1468 |
|
| 1469 | __SCX_NR_CAPS,
|
| 1470 | __SCX_CAP_ALL = BIT_U64(__SCX_NR_CAPS) - 1,
|
| 1471 |
|
| 1472 | SCX_CAP_ENQ_IMMED = BIT_U64(__SCX_CAP_ENQ_IMMED),
|
| 1473 | SCX_CAP_ENQ = BIT_U64(__SCX_CAP_ENQ),
|
| 1474 | SCX_CAP_PREEMPT = BIT_U64(__SCX_CAP_PREEMPT),
|
| 1475 | SCX_CAP_PERF = BIT_U64(__SCX_CAP_PERF),
|
| 1476 |
|
| 1477 |
|
| 1478 | SCX_CAP_BASE = SCX_CAP_ENQ_IMMED,
|
| 1479 |
|
| 1480 |
|
| 1481 | SCX_CAPS_REENQ_ON_LOSS = SCX_CAP_ENQ_IMMED | SCX_CAP_ENQ,
|
| 1482 | };
|
| 1483 |
|
| 1484 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 1485 |
|
| 1486 | #define scx_for_each_cap_bit(cap_bit, caps) \
|
| 1487 | for (u64 __caps = (caps); \
|
| 1488 | __caps && ((cap_bit) = __ffs64(__caps), true); \
|
| 1489 | __caps &= __caps - 1)
|
| 1490 |
|
| 1491 | |
| 1492 | * Sub-cap update notifier.
|
| 1493 | *
|
| 1494 | * ops_cid.sub_caps_updated() notifies sub-scheds when their cap state changes
|
| 1495 | * so they can refresh internal state without polling scx_bpf_sub_caps() per
|
| 1496 | * enqueue.
|
| 1497 | *
|
| 1498 | * Three constraints shape the design:
|
| 1499 | *
|
| 1500 | * 1. Static memory. Deliveries use a fixed-size buffer, both for runtime
|
| 1501 | * efficiency and so notifications can't be lost under memory pressure.
|
| 1502 | *
|
| 1503 | * 2. High-frequency updates. Grant/revoke can mutate caps in bursts, and the
|
| 1504 | * notifier path must absorb that without amplifying it.
|
| 1505 | *
|
| 1506 | * 3. Recursive grant/revoke from the callback. A child receiving a
|
| 1507 | * notification can call grant/revoke on its own children, which can
|
| 1508 | * cascade recursively down its subtree.
|
| 1509 | *
|
| 1510 | * (1) and (2) lead to coalescing into a fixed payload. Each delivery carries a
|
| 1511 | * single (cmask, caps) pair covering every change since the previous one.
|
| 1512 | * Direction (set vs cleared) isn't encoded as it doesn't fit in the fixed-size
|
| 1513 | * summary. The callback queries scx_bpf_sub_caps() for current state. Only one
|
| 1514 | * delivery is in flight per shard. Further changes fold into the same buffer
|
| 1515 | * and ship as the next callback, so a shard's callbacks fire in order.
|
| 1516 | *
|
| 1517 | * (3) leads to deferred delivery. Events accumulate during grant/revoke and are
|
| 1518 | * delivered after the shard lock is released.
|
| 1519 |
|
| 1520 | struct scx_caps_updated {
|
| 1521 | raw_spinlock_t lock;
|
| 1522 | u64 caps;
|
| 1523 | struct scx_cmask *cmask_arena_out;
|
| 1524 | struct list_head node_in_flight;
|
| 1525 |
|
| 1526 | TRAILING_OVERLAP(struct scx_cmask, cmask, bits,
|
| 1527 | u64 _bits[SCX_CMASK_NR_WORDS(SCX_CID_SHARD_MAX_CPUS)];
|
| 1528 | );
|
| 1529 | };
|
| 1530 |
|
| 1531 | struct scx_pshard {
|
| 1532 | raw_spinlock_t lock;
|
| 1533 | struct scx_sched *sch;
|
| 1534 | struct scx_caps_updated caps_updated;
|
| 1535 |
|
| 1536 | |
| 1537 | * Per-cap cmask, inline via TRAILING_OVERLAP so cmask.bits[] overlaps
|
| 1538 | * the trailing _bits[] storage. Access as &caps[i].cmask. See
|
| 1539 | * scx_sched_pcpu->ecaps.
|
| 1540 |
|
| 1541 | TRAILING_OVERLAP(struct scx_cmask, cmask, bits,
|
| 1542 | u64 _bits[SCX_CMASK_NR_WORDS(SCX_CID_SHARD_MAX_CPUS)];
|
| 1543 | ) caps[__SCX_NR_CAPS];
|
| 1544 |
|
| 1545 | |
| 1546 | * Shard geometry captured at alloc. cmask_arena_out's own header is
|
| 1547 | * bpf-writable and the live shard range can change before the
|
| 1548 | * rcu-deferred free, so re-init and size cmask_arena_out from these
|
| 1549 | * trusted copies instead.
|
| 1550 |
|
| 1551 | u32 base;
|
| 1552 | u32 nr_cids;
|
| 1553 | };
|
| 1554 | #endif
|
| 1555 |
|
| 1556 | struct scx_sched {
|
| 1557 | |
| 1558 | * cpu-form and cid-form ops share field offsets up to .priv (verified
|
| 1559 | * by BUILD_BUG_ON in scx_init()). The anonymous union lets the kernel
|
| 1560 | * access either view of the same storage without function-pointer
|
| 1561 | * casts: use .ops for cpu-form and shared fields, .ops_cid for the
|
| 1562 | * callbacks whose cid-form signature differs (set_cmask, enable,
|
| 1563 | * select_cid, cid_online, ...).
|
| 1564 |
|
| 1565 | union {
|
| 1566 | struct sched_ext_ops ops;
|
| 1567 | struct sched_ext_ops_cid ops_cid;
|
| 1568 | };
|
| 1569 | bool is_cid_type;
|
| 1570 | bool dead;
|
| 1571 |
|
| 1572 | |
| 1573 | * Arena map auto-discovered from member progs at struct_ops attach.
|
| 1574 | * cid-form schedulers must use exactly one arena across all member
|
| 1575 | * progs. NULL on cpu-form.
|
| 1576 | *
|
| 1577 | * @arena_pool sub-allocates @arena_map. Each gen_pool chunk is added
|
| 1578 | * at the kernel-side mapping address. @arena_kern_base is the start
|
| 1579 | * of the arena's kern_vm range. See scx_arena_to_kaddr().
|
| 1580 |
|
| 1581 | struct bpf_map *arena_map;
|
| 1582 | struct gen_pool *arena_pool;
|
| 1583 | uintptr_t arena_kern_base;
|
| 1584 |
|
| 1585 | |
| 1586 | * Per-CPU arena cmask the kernel fills from a task's cpumask and hands
|
| 1587 | * to ops_cid.enable() and ops_cid.set_cmask(). The stored pointers are
|
| 1588 | * the kernel addresses.
|
| 1589 |
|
| 1590 | struct scx_cmask * __percpu *set_cmask_scratch;
|
| 1591 | struct scx_cmask *online_cmask;
|
| 1592 |
|
| 1593 | DECLARE_BITMAP(has_op, SCX_OPI_END);
|
| 1594 |
|
| 1595 | |
| 1596 | * Dispatch queues.
|
| 1597 | *
|
| 1598 | * The global DSQ (%SCX_DSQ_GLOBAL) is split per-node for scalability.
|
| 1599 | * This is to avoid live-locking in bypass mode where all tasks are
|
| 1600 | * dispatched to %SCX_DSQ_GLOBAL and all CPUs consume from it. If
|
| 1601 | * per-node split isn't sufficient, it can be further split.
|
| 1602 |
|
| 1603 | struct rhashtable dsq_hash;
|
| 1604 | struct scx_sched_pnode **pnode;
|
| 1605 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 1606 | struct scx_pshard **pshard;
|
| 1607 | #endif
|
| 1608 | struct scx_sched_pcpu __percpu *pcpu;
|
| 1609 |
|
| 1610 | u64 slice_dfl;
|
| 1611 | u64 bypass_timestamp;
|
| 1612 | s32 bypass_depth;
|
| 1613 |
|
| 1614 |
|
| 1615 | unsigned long bypass_dsp_claim;
|
| 1616 | atomic_t bypass_dsp_enable_depth;
|
| 1617 |
|
| 1618 | bool aborting;
|
| 1619 | bool dump_disabled;
|
| 1620 | u32 dsp_max_batch;
|
| 1621 | s32 level;
|
| 1622 |
|
| 1623 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 1624 | |
| 1625 | * pshard[] size captured at enable for the async RCU free path -
|
| 1626 | * scx_nr_cid_shards may be rewritten by a later enable's
|
| 1627 | * scx_cid_publish_tables() before free runs. While sch is active, use
|
| 1628 | * the global.
|
| 1629 |
|
| 1630 | u32 nr_pshards;
|
| 1631 | #endif
|
| 1632 |
|
| 1633 | |
| 1634 | * Updates to the following warned bitfields can race causing RMW issues
|
| 1635 | * but it doesn't really matter.
|
| 1636 |
|
| 1637 | bool warned_zero_slice:1;
|
| 1638 | bool warned_unassoc_progs:1;
|
| 1639 |
|
| 1640 | struct list_head all;
|
| 1641 |
|
| 1642 |
|
| 1643 | u64 id;
|
| 1644 |
|
| 1645 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 1646 | struct rhash_head hash_node;
|
| 1647 |
|
| 1648 | struct list_head children;
|
| 1649 | struct list_head sibling;
|
| 1650 | struct cgroup *cgrp;
|
| 1651 | char *cgrp_path;
|
| 1652 | struct kset *sub_kset;
|
| 1653 |
|
| 1654 | bool linked;
|
| 1655 | bool sub_attached;
|
| 1656 | #endif
|
| 1657 |
|
| 1658 | |
| 1659 | * The maximum amount of time in jiffies that a task may be runnable
|
| 1660 | * without being scheduled on a CPU. If this timeout is exceeded, it
|
| 1661 | * will trigger scx_error().
|
| 1662 |
|
| 1663 | unsigned long watchdog_timeout;
|
| 1664 |
|
| 1665 | atomic_t exit_kind;
|
| 1666 | struct scx_exit_info *exit_info;
|
| 1667 |
|
| 1668 | struct kobject kobj;
|
| 1669 |
|
| 1670 | struct kthread_worker *helper;
|
| 1671 | struct irq_work disable_irq_work;
|
| 1672 | struct kthread_work disable_work;
|
| 1673 | struct irq_work propagate_exit_irq_work;
|
| 1674 | struct timer_list bypass_lb_timer;
|
| 1675 | cpumask_var_t bypass_lb_donee_cpumask;
|
| 1676 | cpumask_var_t bypass_lb_resched_cpumask;
|
| 1677 | cpumask_var_t stall_cpus;
|
| 1678 | struct rcu_work rcu_work;
|
| 1679 |
|
| 1680 |
|
| 1681 | struct scx_sched *ancestors[];
|
| 1682 | };
|
| 1683 |
|
| 1684 | |
| 1685 | * scx_arena_to_kaddr - Translate a BPF-arena pointer to its kernel address
|
| 1686 | * @sch: scheduler whose arena hosts @bpf_ptr
|
| 1687 | * @bpf_ptr: BPF-arena pointer, only the low 32 bits are used
|
| 1688 | *
|
| 1689 | * The (u32) cast normalizes any input into the arena's 4 GiB kern_vm range,
|
| 1690 | * which combined with scratch-page fault recovery makes the returned pointer
|
| 1691 | * safe to dereference up to GUARD_SZ / 2 past the intended object. Accesses
|
| 1692 | * larger than GUARD_SZ / 2 must be explicitly bounds-checked.
|
| 1693 |
|
| 1694 | static inline void *scx_arena_to_kaddr(struct scx_sched *sch, const void *bpf_ptr)
|
| 1695 | {
|
| 1696 | return (void *)(sch->arena_kern_base + (u32)(uintptr_t)bpf_ptr);
|
| 1697 | }
|
| 1698 |
|
| 1699 | |
| 1700 | * scx_kaddr_to_arena - Translate a kernel arena address to the BPF form
|
| 1701 | * @sch: scheduler whose arena hosts @kaddr
|
| 1702 | * @kaddr: kernel address inside @sch's arena
|
| 1703 | *
|
| 1704 | * __arena callback arguments need no translation. Addresses handed to BPF any
|
| 1705 | * other way, such as struct fields and kfunc return values, go through this.
|
| 1706 |
|
| 1707 | static inline uintptr_t scx_kaddr_to_arena(struct scx_sched *sch, const void *kaddr)
|
| 1708 | {
|
| 1709 | return (uintptr_t)kaddr - sch->arena_kern_base;
|
| 1710 | }
|
| 1711 |
|
| 1712 | enum scx_wake_flags {
|
| 1713 |
|
| 1714 | SCX_WAKE_FORK = WF_FORK,
|
| 1715 | SCX_WAKE_TTWU = WF_TTWU,
|
| 1716 | SCX_WAKE_SYNC = WF_SYNC,
|
| 1717 | };
|
| 1718 |
|
| 1719 | enum scx_enq_flags {
|
| 1720 |
|
| 1721 | SCX_ENQ_WAKEUP = ENQUEUE_WAKEUP,
|
| 1722 | SCX_ENQ_HEAD = ENQUEUE_HEAD,
|
| 1723 | SCX_ENQ_CPU_SELECTED = ENQUEUE_RQ_SELECTED,
|
| 1724 |
|
| 1725 |
|
| 1726 |
|
| 1727 | |
| 1728 | * Set the following to trigger preemption when calling
|
| 1729 | * scx_bpf_dsq_insert() with a local dsq as the target. The slice of the
|
| 1730 | * current task is cleared to zero and the CPU is kicked into the
|
| 1731 | * scheduling path. Implies %SCX_ENQ_HEAD.
|
| 1732 |
|
| 1733 | SCX_ENQ_PREEMPT = 1LLU << 32,
|
| 1734 |
|
| 1735 | |
| 1736 | * Only allowed on local DSQs. Guarantees that the task either gets
|
| 1737 | * on the CPU immediately and stays on it, or gets reenqueued back
|
| 1738 | * to the BPF scheduler. It will never linger on a local DSQ or be
|
| 1739 | * silently put back after preemption.
|
| 1740 | *
|
| 1741 | * The protection persists until the next fresh enqueue - it
|
| 1742 | * survives SAVE/RESTORE cycles, slice extensions and preemption.
|
| 1743 | * If the task can't stay on the CPU for any reason, it gets
|
| 1744 | * reenqueued back to the BPF scheduler.
|
| 1745 | *
|
| 1746 | * Exiting and migration-disabled tasks bypass ops.enqueue() and
|
| 1747 | * are placed directly on a local DSQ without IMMED protection
|
| 1748 | * unless %SCX_OPS_ENQ_EXITING and %SCX_OPS_ENQ_MIGRATION_DISABLED
|
| 1749 | * are set respectively.
|
| 1750 |
|
| 1751 | SCX_ENQ_IMMED = 1LLU << 33,
|
| 1752 |
|
| 1753 | |
| 1754 | * Only allowed on local DSQs. If the insert lacks the caps for the
|
| 1755 | * target cid, divert the task to the CPU's rescue path instead of
|
| 1756 | * rejecting and reenqueueing, e.g. when the task's affinity is
|
| 1757 | * restricted to cids the scheduler doesn't hold. The kernel runs
|
| 1758 | * rescued tasks on the target CPU. Rescue execution is guaranteed to
|
| 1759 | * make forward progress and is bandwidth-limited, see the
|
| 1760 | * rescue_bandwidth_ppt and rescue_quantum_us ops fields.
|
| 1761 |
|
| 1762 | SCX_ENQ_RESCUE = 1LLU << 34,
|
| 1763 |
|
| 1764 | |
| 1765 | * The task being enqueued was previously enqueued on a DSQ, but was
|
| 1766 | * removed and is being re-enqueued. See SCX_TASK_REENQ_* flags to find
|
| 1767 | * out why a given task is being reenqueued.
|
| 1768 |
|
| 1769 | SCX_ENQ_REENQ = 1LLU << 40,
|
| 1770 |
|
| 1771 | |
| 1772 | * The task being enqueued is the only task available for the cpu. By
|
| 1773 | * default, ext core keeps executing such tasks but when
|
| 1774 | * %SCX_OPS_ENQ_LAST is specified, they're ops.enqueue()'d with the
|
| 1775 | * %SCX_ENQ_LAST flag set.
|
| 1776 | *
|
| 1777 | * If the task is queued on the local DSQ of the CPU it was running on,
|
| 1778 | * it continues to run. Otherwise, the CPU goes idle. A scheduler that
|
| 1779 | * wants a full dispatch cycle on the CPU should kick it.
|
| 1780 |
|
| 1781 | SCX_ENQ_LAST = 1LLU << 41,
|
| 1782 |
|
| 1783 |
|
| 1784 | __SCX_ENQ_INTERNAL_MASK = 0xffLLU << 56,
|
| 1785 |
|
| 1786 | SCX_ENQ_CLEAR_OPSS = 1LLU << 56,
|
| 1787 | SCX_ENQ_DSQ_PRIQ = 1LLU << 57,
|
| 1788 | SCX_ENQ_NESTED = 1LLU << 58,
|
| 1789 | SCX_ENQ_GDSQ_FALLBACK = 1LLU << 59,
|
| 1790 | SCX_ENQ_IGNORE_CAPS = 1LLU << 60,
|
| 1791 | SCX_ENQ_APPLY_SLICE = 1LLU << 61,
|
| 1792 | SCX_ENQ_SLICE_DFL = 1LLU << 62,
|
| 1793 | };
|
| 1794 |
|
| 1795 | enum scx_deq_flags {
|
| 1796 |
|
| 1797 | SCX_DEQ_SLEEP = DEQUEUE_SLEEP,
|
| 1798 |
|
| 1799 |
|
| 1800 |
|
| 1801 | |
| 1802 | * The generic core-sched layer decided to execute the task even though
|
| 1803 | * it hasn't been dispatched yet. Dequeue from the BPF side.
|
| 1804 |
|
| 1805 | SCX_DEQ_CORE_SCHED_EXEC = 1LLU << 32,
|
| 1806 |
|
| 1807 | |
| 1808 | * The task is being dequeued due to a property change (e.g.,
|
| 1809 | * sched_setaffinity(), sched_setscheduler(), set_user_nice(),
|
| 1810 | * etc.).
|
| 1811 |
|
| 1812 | SCX_DEQ_SCHED_CHANGE = 1LLU << 33,
|
| 1813 | };
|
| 1814 |
|
| 1815 | enum scx_reenq_flags {
|
| 1816 |
|
| 1817 | SCX_REENQ_ANY = 1LLU << 0,
|
| 1818 |
|
| 1819 |
|
| 1820 | SCX_REENQ_CAP_REVOKE = 1LLU << 1,
|
| 1821 |
|
| 1822 | __SCX_REENQ_FILTER_MASK = 0xffffLLU,
|
| 1823 |
|
| 1824 | __SCX_REENQ_USER_MASK = SCX_REENQ_ANY,
|
| 1825 |
|
| 1826 |
|
| 1827 | SCX_REENQ_TSR_RQ_OPEN = 1LLU << 32,
|
| 1828 | SCX_REENQ_TSR_NOT_FIRST = 1LLU << 33,
|
| 1829 |
|
| 1830 | __SCX_REENQ_TSR_MASK = 0xfLLU << 32,
|
| 1831 | };
|
| 1832 |
|
| 1833 | enum scx_pick_idle_cpu_flags {
|
| 1834 | SCX_PICK_IDLE_CORE = 1LLU << 0,
|
| 1835 | SCX_PICK_IDLE_IN_NODE = 1LLU << 1,
|
| 1836 | };
|
| 1837 |
|
| 1838 | enum scx_kick_flags {
|
| 1839 | |
| 1840 | * Kick the target CPU if idle. Guarantees that the target CPU goes
|
| 1841 | * through at least one full scheduling cycle before going idle. If the
|
| 1842 | * target CPU can be determined to be currently not idle and going to go
|
| 1843 | * through a scheduling cycle before going idle, noop.
|
| 1844 |
|
| 1845 | SCX_KICK_IDLE = 1LLU << 0,
|
| 1846 |
|
| 1847 | |
| 1848 | * Preempt the current task and execute the dispatch path. If the
|
| 1849 | * current task of the target CPU is an SCX task, its ->scx.slice is
|
| 1850 | * cleared to zero before the scheduling path is invoked so that the
|
| 1851 | * task expires and the dispatch path is invoked.
|
| 1852 |
|
| 1853 | SCX_KICK_PREEMPT = 1LLU << 1,
|
| 1854 |
|
| 1855 | |
| 1856 | * The scx_bpf_kick_cpu() call will return after the current SCX task of
|
| 1857 | * the target CPU switches out. This can be used to implement e.g. core
|
| 1858 | * scheduling. This has no effect if the current task on the target CPU
|
| 1859 | * is not on SCX.
|
| 1860 |
|
| 1861 | SCX_KICK_WAIT = 1LLU << 2,
|
| 1862 | };
|
| 1863 |
|
| 1864 | enum scx_tg_flags {
|
| 1865 | SCX_TG_ONLINE = 1U << 0,
|
| 1866 | SCX_TG_INITED = 1U << 1,
|
| 1867 | SCX_TG_SUB_INIT = 1U << 2,
|
| 1868 | };
|
| 1869 |
|
| 1870 | enum scx_enable_state {
|
| 1871 | SCX_ENABLING,
|
| 1872 | SCX_ENABLED,
|
| 1873 | SCX_DISABLING,
|
| 1874 | SCX_DISABLED,
|
| 1875 | };
|
| 1876 |
|
| 1877 | static const char *scx_enable_state_str[] = {
|
| 1878 | [SCX_ENABLING] = "enabling",
|
| 1879 | [SCX_ENABLED] = "enabled",
|
| 1880 | [SCX_DISABLING] = "disabling",
|
| 1881 | [SCX_DISABLED] = "disabled",
|
| 1882 | };
|
| 1883 |
|
| 1884 | |
| 1885 | * Task Ownership State Machine (sched_ext_entity->ops_state)
|
| 1886 | *
|
| 1887 | * The sched_ext core uses this state machine to track task ownership
|
| 1888 | * between the SCX core and the BPF scheduler. This allows the BPF
|
| 1889 | * scheduler to dispatch tasks without strict ordering requirements, while
|
| 1890 | * the SCX core safely rejects invalid dispatches.
|
| 1891 | *
|
| 1892 | * State Transitions
|
| 1893 | *
|
| 1894 | * .------------> NONE (owned by SCX core)
|
| 1895 | * | | ^
|
| 1896 | * | enqueue | | direct dispatch
|
| 1897 | * | v |
|
| 1898 | * | QUEUEING -------'
|
| 1899 | * | |
|
| 1900 | * | enqueue |
|
| 1901 | * | completes |
|
| 1902 | * | v
|
| 1903 | * | QUEUED (owned by BPF scheduler)
|
| 1904 | * | |
|
| 1905 | * | dispatch |
|
| 1906 | * | |
|
| 1907 | * | v
|
| 1908 | * | DISPATCHING
|
| 1909 | * | |
|
| 1910 | * | dispatch |
|
| 1911 | * | completes |
|
| 1912 | * `---------------'
|
| 1913 | *
|
| 1914 | * State Descriptions
|
| 1915 | *
|
| 1916 | * - %SCX_OPSS_NONE:
|
| 1917 | * Task is owned by the SCX core. It's either on a run queue, running,
|
| 1918 | * or being manipulated by the core scheduler. The BPF scheduler has no
|
| 1919 | * claim on this task.
|
| 1920 | *
|
| 1921 | * - %SCX_OPSS_QUEUEING:
|
| 1922 | * Transitional state while transferring a task from the SCX core to
|
| 1923 | * the BPF scheduler. The task's rq lock is held during this state.
|
| 1924 | * Since QUEUEING is both entered and exited under the rq lock, dequeue
|
| 1925 | * can never observe this state (it would be a BUG). When finishing a
|
| 1926 | * dispatch, if the task is still in %SCX_OPSS_QUEUEING the completion
|
| 1927 | * path busy-waits for it to leave this state (via wait_ops_state())
|
| 1928 | * before retrying.
|
| 1929 | *
|
| 1930 | * - %SCX_OPSS_QUEUED:
|
| 1931 | * Task is owned by the BPF scheduler. It's on a DSQ (dispatch queue)
|
| 1932 | * and the BPF scheduler is responsible for dispatching it. A QSEQ
|
| 1933 | * (queue sequence number) is embedded in this state to detect
|
| 1934 | * dispatch/dequeue races: if a task is dequeued and re-enqueued, the
|
| 1935 | * QSEQ changes and any in-flight dispatch operations targeting the old
|
| 1936 | * QSEQ are safely ignored.
|
| 1937 | *
|
| 1938 | * - %SCX_OPSS_DISPATCHING:
|
| 1939 | * Transitional state while transferring a task from the BPF scheduler
|
| 1940 | * back to the SCX core. This state indicates the BPF scheduler has
|
| 1941 | * selected the task for execution. When dequeue needs to take the task
|
| 1942 | * off a DSQ and it is still in %SCX_OPSS_DISPATCHING, the dequeue path
|
| 1943 | * busy-waits for it to leave this state (via wait_ops_state()) before
|
| 1944 | * proceeding. Exits to %SCX_OPSS_NONE when dispatch completes.
|
| 1945 | *
|
| 1946 | * Memory Ordering
|
| 1947 | *
|
| 1948 | * Transitions out of %SCX_OPSS_QUEUEING and %SCX_OPSS_DISPATCHING into
|
| 1949 | * %SCX_OPSS_NONE or %SCX_OPSS_QUEUED must use atomic_long_set_release()
|
| 1950 | * and waiters must use atomic_long_read_acquire(). This ensures proper
|
| 1951 | * synchronization between concurrent operations.
|
| 1952 | *
|
| 1953 | * Cross-CPU Task Migration
|
| 1954 | *
|
| 1955 | * When moving a task in the %SCX_OPSS_DISPATCHING state, we can't simply
|
| 1956 | * grab the target CPU's rq lock because a concurrent dequeue might be
|
| 1957 | * waiting on %SCX_OPSS_DISPATCHING while holding the source rq lock
|
| 1958 | * (deadlock).
|
| 1959 | *
|
| 1960 | * The sched_ext core uses a "lock dancing" protocol coordinated by
|
| 1961 | * p->scx.holding_cpu. When moving a task to a different rq:
|
| 1962 | *
|
| 1963 | * 1. Set p->scx.holding_cpu to the current CPU
|
| 1964 | * 2. Set task state to %SCX_OPSS_NONE; dequeue waits while DISPATCHING
|
| 1965 | * is set, so clearing DISPATCHING first prevents the circular wait
|
| 1966 | * (safe to lock the rq we need)
|
| 1967 | * 3. Unlock the current CPU's rq
|
| 1968 | * 4. Lock src_rq (where the task currently lives)
|
| 1969 | * 5. Verify p->scx.holding_cpu == current CPU, if not, dequeue won the
|
| 1970 | * race (dequeue clears holding_cpu to -1 when it takes the task), in
|
| 1971 | * this case migration is aborted
|
| 1972 | * 6. If src_rq == dst_rq: clear holding_cpu and enqueue directly
|
| 1973 | * into dst_rq's local DSQ (no lock swap needed)
|
| 1974 | * 7. Otherwise, verify under src_rq lock that the task can be moved to dst_rq
|
| 1975 | * (CPU affinity, migration_disabled, etc.). If not, clear holding_cpu,
|
| 1976 | * leave the task on src_rq, and enqueue it on the fallback DSQ.
|
| 1977 | * 8. Otherwise (i.e. if the task can be moved to dst_rq), call
|
| 1978 | * move_remote_task_to_local_dsq(), which releases src_rq, locks dst_rq,
|
| 1979 | * and performs the deactivate/activate migration cycle
|
| 1980 | * (dst_rq is held on return)
|
| 1981 | * 9. Unlock dst_rq and re-lock the current CPU's rq to restore
|
| 1982 | * the lock state expected by the caller
|
| 1983 | *
|
| 1984 | * If any verification fails, abort the migration.
|
| 1985 | *
|
| 1986 | * This state tracking allows the BPF scheduler to try to dispatch any task
|
| 1987 | * at any time regardless of its state. The SCX core can safely
|
| 1988 | * reject/ignore invalid dispatches, simplifying the BPF scheduler
|
| 1989 | * implementation.
|
| 1990 |
|
| 1991 | enum scx_ops_state {
|
| 1992 | SCX_OPSS_NONE,
|
| 1993 | SCX_OPSS_QUEUEING,
|
| 1994 | SCX_OPSS_QUEUED,
|
| 1995 | SCX_OPSS_DISPATCHING,
|
| 1996 |
|
| 1997 | |
| 1998 | * QSEQ brands each QUEUED instance so that, when dispatch races
|
| 1999 | * dequeue/requeue, the dispatcher can tell whether it still has a claim
|
| 2000 | * on the task being dispatched.
|
| 2001 | *
|
| 2002 | * QSEQ is generated from the per-task p->scx.ops_qseq counter so that
|
| 2003 | * it doesn't repeat across QUEUED instances of the same task even if
|
| 2004 | * the task moves between rqs. 0 is never used as a valid QSEQ since
|
| 2005 | * NONE and DISPATCHING map to this value.
|
| 2006 | *
|
| 2007 | * As some 32bit archs can't do 64bit store_release/load_acquire,
|
| 2008 | * p->scx.ops_state is atomic_long_t which leaves 30 bits for QSEQ on
|
| 2009 | * 32bit machines. The dispatch race window QSEQ protects is very narrow
|
| 2010 | * and runs with IRQ disabled. 30 bits should be sufficient.
|
| 2011 |
|
| 2012 | SCX_OPSS_QSEQ_SHIFT = 2,
|
| 2013 | };
|
| 2014 |
|
| 2015 |
|
| 2016 | #define SCX_OPSS_STATE_MASK ((1LU << SCX_OPSS_QSEQ_SHIFT) - 1)
|
| 2017 | #define SCX_OPSS_QSEQ_MASK (~SCX_OPSS_STATE_MASK)
|
| 2018 |
|
| 2019 | |
| 2020 | * SCX task iterator.
|
| 2021 |
|
| 2022 | struct scx_task_iter {
|
| 2023 | struct sched_ext_entity cursor;
|
| 2024 | struct task_struct *locked_task;
|
| 2025 | struct rq *rq;
|
| 2026 | struct rq_flags rf;
|
| 2027 | u32 cnt;
|
| 2028 | bool list_locked;
|
| 2029 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 2030 | struct cgroup *cgrp;
|
| 2031 | struct cgroup_subsys_state *css_pos;
|
| 2032 | struct css_task_iter css_iter;
|
| 2033 | #endif
|
| 2034 | };
|
| 2035 |
|
| 2036 | |
| 2037 | * scx_enable() is offloaded to a dedicated system-wide RT kthread to avoid
|
| 2038 | * starvation. During the READY -> ENABLED task switching loop, the calling
|
| 2039 | * thread's sched_class gets switched from fair to ext. As fair has higher
|
| 2040 | * priority than ext, the calling thread can be indefinitely starved under
|
| 2041 | * fair-class saturation, leading to a system hang.
|
| 2042 |
|
| 2043 | struct scx_enable_cmd {
|
| 2044 | struct kthread_work work;
|
| 2045 | union {
|
| 2046 | struct sched_ext_ops *ops;
|
| 2047 | struct sched_ext_ops_cid *ops_cid;
|
| 2048 | };
|
| 2049 | bool is_cid_type;
|
| 2050 | struct bpf_map *arena_map;
|
| 2051 | int ret;
|
| 2052 | };
|
| 2053 |
|
| 2054 |
|
| 2055 | struct scx_bstr_buf {
|
| 2056 | u64 data[MAX_BPRINTF_VARARGS];
|
| 2057 | char line[SCX_EXIT_MSG_LEN];
|
| 2058 | };
|
| 2059 |
|
| 2060 |
|
| 2061 | #define DECLARE_SCX_COMPAT_MARKER(func) \
|
| 2062 | extern void scx_compat_marker_##func(void)
|
| 2063 |
|
| 2064 | |
| 2065 | * DEFINE_SCX_COMPAT_MARKER() - define a userspace capability marker
|
| 2066 | * @func: marker suffix; the defined symbol is scx_compat_marker_@func
|
| 2067 | *
|
| 2068 | * Emit an empty, callerless function that is retained in the kernel's BTF.
|
| 2069 | * Its presence is part of the kernel<->userspace contract: userspace probes
|
| 2070 | * scx_compat_marker_@func (e.g. via BTF) to detect that this kernel supports
|
| 2071 | * the corresponding feature.
|
| 2072 | *
|
| 2073 | * The leading declaration suppresses the missing-prototype warning; the
|
| 2074 | * trailing declaration consumes the semicolon at the use site.
|
| 2075 |
|
| 2076 | #define DEFINE_SCX_COMPAT_MARKER(func) \
|
| 2077 | DECLARE_SCX_COMPAT_MARKER(func); \
|
| 2078 | __used __retain void scx_compat_marker_##func(void) {} \
|
| 2079 | DECLARE_SCX_COMPAT_MARKER(func)
|
| 2080 |
|
| 2081 | extern struct scx_sched __rcu *scx_root;
|
| 2082 | DECLARE_PER_CPU(struct rq *, scx_locked_rq_state);
|
| 2083 |
|
| 2084 | |
| 2085 | * True when the currently loaded scheduler hierarchy is cid-form. All scheds
|
| 2086 | * in a hierarchy share one form, so this single key tells callsites which
|
| 2087 | * view to use without per-sch dereferences. Use scx_is_cid_type() to test.
|
| 2088 |
|
| 2089 | DECLARE_STATIC_KEY_FALSE(__scx_is_cid_type);
|
| 2090 |
|
| 2091 | int scx_kfunc_context_filter(const struct bpf_prog *prog, u32 kfunc_id);
|
| 2092 |
|
| 2093 | bool scx_cpu_valid(struct scx_sched *sch, s32 cpu, const char *where);
|
| 2094 |
|
| 2095 | __printf(5, 0) bool scx_vexit(struct scx_sched *sch, enum scx_exit_kind kind,
|
| 2096 | s64 exit_code, s32 exit_cpu, const char *fmt,
|
| 2097 | va_list args);
|
| 2098 | __printf(5, 6) bool __scx_exit(struct scx_sched *sch, enum scx_exit_kind kind,
|
| 2099 | s64 exit_code, s32 exit_cpu, const char *fmt, ...);
|
| 2100 |
|
| 2101 | u32 scx_get_task_state(const struct task_struct *p);
|
| 2102 | void scx_set_task_state(struct task_struct *p, u32 state);
|
| 2103 | void scx_task_iter_start(struct scx_task_iter *iter, struct cgroup *cgrp);
|
| 2104 | void scx_task_iter_unlock(struct scx_task_iter *iter);
|
| 2105 | void scx_task_iter_stop(struct scx_task_iter *iter);
|
| 2106 | struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter);
|
| 2107 | bool scx_set_task_slice(struct task_struct *p, u64 slice);
|
| 2108 | void scx_task_slice_ended(struct rq *rq, struct task_struct *p);
|
| 2109 | void scx_task_unlink_from_dsq(struct task_struct *p, struct scx_dispatch_q *dsq);
|
| 2110 | void scx_dispatch_dequeue(struct rq *rq, struct task_struct *p);
|
| 2111 | void scx_do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
|
| 2112 | int sticky_cpu);
|
| 2113 | void scx_move_local_task_to_local_dsq(struct scx_sched *sch, struct task_struct *p,
|
| 2114 | u64 enq_flags, struct rq *dst_rq);
|
| 2115 | bool scx_consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
|
| 2116 | struct scx_dispatch_q *dsq, u64 enq_flags);
|
| 2117 | bool scx_consume_global_dsq(struct scx_sched *sch, struct rq *rq);
|
| 2118 | bool scx_rq_online(struct rq *rq);
|
| 2119 | void scx_flush_dispatch_buf(struct scx_sched *sch, struct rq *rq);
|
| 2120 | s32 scx_init_dsq(struct scx_dispatch_q *dsq, u64 dsq_id, struct scx_sched *sch);
|
| 2121 | __printf(2, 3) void scx_dump_line(struct seq_buf *s, const char *fmt, ...);
|
| 2122 | void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags);
|
| 2123 | u64 __scx_bpf_now(struct rq *rq);
|
| 2124 | void schedule_dsq_reenq(struct scx_sched *sch, struct scx_dispatch_q *dsq,
|
| 2125 | u64 reenq_flags, struct rq *locked_rq);
|
| 2126 | void scx_wait_dispatching(struct task_struct *p);
|
| 2127 | int __scx_init_task(struct scx_sched *sch, struct task_struct *p,
|
| 2128 | struct cgroup *cgrp, bool fork);
|
| 2129 | void scx_enable_task(struct scx_sched *sch, struct task_struct *p);
|
| 2130 | void __scx_disable_and_exit_task(struct scx_sched *sch, struct task_struct *p);
|
| 2131 | void scx_sub_init_cancel_task(struct scx_sched *sch, struct task_struct *p);
|
| 2132 | void scx_disable_and_exit_task(struct scx_sched *sch, struct task_struct *p);
|
| 2133 | #if defined(CONFIG_EXT_GROUP_SCHED) || defined(CONFIG_EXT_SUB_SCHED)
|
| 2134 | void scx_cgroup_lock(void);
|
| 2135 | void scx_cgroup_unlock(void);
|
| 2136 | #endif
|
| 2137 | s32 scx_alloc_kern_arena_objs(struct scx_sched *sch);
|
| 2138 | void scx_disable_bypass_dsp(struct scx_sched *sch);
|
| 2139 | void scx_bypass(struct scx_sched *sch, bool bypass);
|
| 2140 | s32 scx_link_sched(struct scx_sched *sch);
|
| 2141 | void scx_unlink_sched(struct scx_sched *sch);
|
| 2142 | void scx_disable_dump(struct scx_sched *sch);
|
| 2143 | void scx_log_sched_disable(struct scx_sched *sch);
|
| 2144 | void scx_flush_disable_work(struct scx_sched *sch);
|
| 2145 | struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
|
| 2146 | struct cgroup *cgrp,
|
| 2147 | struct scx_sched *parent);
|
| 2148 | int scx_validate_ops(struct scx_sched *sch, const struct sched_ext_ops *ops);
|
| 2149 | int scx_sched_sysfs_add(struct scx_sched *sch);
|
| 2150 | bool scx_is_descendant(struct scx_sched *sch, struct scx_sched *ancestor);
|
| 2151 | __printf(5, 0) bool scx_exit_bstr(struct scx_sched *sch, enum scx_exit_kind kind,
|
| 2152 | s64 exit_code, struct scx_sched *fmt_blame,
|
| 2153 | char *fmt, unsigned long long *data, u32 data__sz);
|
| 2154 |
|
| 2155 | extern raw_spinlock_t scx_sched_lock;
|
| 2156 | extern struct mutex scx_enable_mutex;
|
| 2157 | extern struct percpu_rw_semaphore scx_fork_rwsem;
|
| 2158 | extern bool scx_cgroup_enabled;
|
| 2159 | extern struct list_head scx_sched_all;
|
| 2160 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 2161 | extern const struct rhashtable_params scx_sched_hash_params;
|
| 2162 | extern struct rhashtable scx_sched_hash;
|
| 2163 | extern struct scx_sched *scx_enabling_sub_sched;
|
| 2164 | #endif
|
| 2165 |
|
| 2166 | #define scx_exit(sch, kind, exit_code, fmt, args...) \
|
| 2167 | __scx_exit(sch, kind, exit_code, raw_smp_processor_id(), fmt, ##args)
|
| 2168 | #define scx_error(sch, fmt, args...) \
|
| 2169 | scx_exit((sch), SCX_EXIT_ERROR, 0, fmt, ##args)
|
| 2170 |
|
| 2171 | |
| 2172 | * scx_root_protected_live - Root sched for paths that only run while live
|
| 2173 | *
|
| 2174 | * scx_root is published before the scheduler goes live and cleared only after
|
| 2175 | * it is fully drained, so a path that only executes while the scheduler is live
|
| 2176 | * can never race an update. Return the root sched with a plain load, never
|
| 2177 | * %NULL.
|
| 2178 |
|
| 2179 | static inline struct scx_sched *scx_root_protected_live(void)
|
| 2180 | {
|
| 2181 | return rcu_dereference_protected(scx_root, true);
|
| 2182 | }
|
| 2183 |
|
| 2184 | |
| 2185 | * scx_root_protected - Root sched for contexts that exclude its updates
|
| 2186 | *
|
| 2187 | * Both scx_root updates run under the locks checked below, so holding one
|
| 2188 | * excludes them. Return the root sched with a plain load, %NULL if no scheduler
|
| 2189 | * is loaded.
|
| 2190 |
|
| 2191 | static inline struct scx_sched *scx_root_protected(void)
|
| 2192 | {
|
| 2193 | return rcu_dereference_protected(scx_root,
|
| 2194 | lockdep_is_cpus_held() ||
|
| 2195 | lockdep_is_held(&scx_enable_mutex));
|
| 2196 | }
|
| 2197 |
|
| 2198 | static inline struct scx_dispatch_q *scx_bypass_dsq(struct scx_sched *sch, s32 cpu)
|
| 2199 | {
|
| 2200 | return &per_cpu_ptr(sch->pcpu, cpu)->bypass_dsq;
|
| 2201 | }
|
| 2202 |
|
| 2203 | |
| 2204 | * scx_bypass_dsp_enabled - Check if bypass dispatch path is enabled
|
| 2205 | * @sch: scheduler to check
|
| 2206 | *
|
| 2207 | * When a descendant scheduler enters bypass mode, bypassed tasks are scheduled
|
| 2208 | * by the nearest non-bypassing ancestor, or the root scheduler if all ancestors
|
| 2209 | * are bypassing. In the former case, the ancestor is not itself bypassing but
|
| 2210 | * its bypass DSQs will be populated with bypassed tasks from descendants. Thus,
|
| 2211 | * the ancestor's bypass dispatch path must be active even though its own
|
| 2212 | * bypass_depth remains zero.
|
| 2213 | *
|
| 2214 | * This function checks bypass_dsp_enable_depth which is managed separately from
|
| 2215 | * bypass_depth to enable this decoupling. See enable_bypass_dsp() and
|
| 2216 | * scx_disable_bypass_dsp().
|
| 2217 |
|
| 2218 | static inline bool scx_bypass_dsp_enabled(struct scx_sched *sch)
|
| 2219 | {
|
| 2220 | return unlikely(atomic_read(&sch->bypass_dsp_enable_depth));
|
| 2221 | }
|
| 2222 |
|
| 2223 | |
| 2224 | * scx_ops_sanitize_err - Sanitize a -errno value
|
| 2225 | * @sch: scx_sched to error out on error
|
| 2226 | * @ops_name: operation to blame on failure
|
| 2227 | * @err: -errno value to sanitize
|
| 2228 | *
|
| 2229 | * Verify @err is a valid -errno. If not, trigger scx_error() and return
|
| 2230 | * -%EPROTO. This is necessary because returning a rogue -errno up the chain can
|
| 2231 | * cause misbehaviors. For an example, a large negative return from
|
| 2232 | * ops.init_task() triggers an oops when passed up the call chain because the
|
| 2233 | * value fails IS_ERR() test after being encoded with ERR_PTR() and then is
|
| 2234 | * handled as a pointer.
|
| 2235 |
|
| 2236 | static inline int scx_ops_sanitize_err(struct scx_sched *sch, const char *ops_name, s32 err)
|
| 2237 | {
|
| 2238 | if (err < 0 && err >= -MAX_ERRNO)
|
| 2239 | return err;
|
| 2240 |
|
| 2241 | scx_error(sch, "ops.%s() returned an invalid errno %d", ops_name, err);
|
| 2242 | return -EPROTO;
|
| 2243 | }
|
| 2244 |
|
| 2245 | static inline void scx_schedule_reenq_local(struct rq *rq, u64 reenq_flags)
|
| 2246 | {
|
| 2247 | struct scx_sched *root = rcu_dereference_sched(scx_root);
|
| 2248 |
|
| 2249 | if (WARN_ON_ONCE(!root))
|
| 2250 | return;
|
| 2251 |
|
| 2252 | schedule_dsq_reenq(root, &rq->scx.local_dsq, reenq_flags, rq);
|
| 2253 | }
|
| 2254 |
|
| 2255 | |
| 2256 | * Return the rq currently locked from an scx callback, or NULL if no rq is
|
| 2257 | * locked.
|
| 2258 |
|
| 2259 | static inline struct rq *scx_locked_rq(void)
|
| 2260 | {
|
| 2261 | return __this_cpu_read(scx_locked_rq_state);
|
| 2262 | }
|
| 2263 |
|
| 2264 | static inline void update_locked_rq(struct rq *rq)
|
| 2265 | {
|
| 2266 | |
| 2267 | * Check whether @rq is actually locked. This can help expose bugs
|
| 2268 | * or incorrect assumptions about the context in which a kfunc or
|
| 2269 | * callback is executed.
|
| 2270 |
|
| 2271 | if (rq)
|
| 2272 | lockdep_assert_rq_held(rq);
|
| 2273 | __this_cpu_write(scx_locked_rq_state, rq);
|
| 2274 | }
|
| 2275 |
|
| 2276 | #define SCX_HAS_OP(sch, op) test_bit(SCX_OP_IDX(op), (sch)->has_op)
|
| 2277 |
|
| 2278 | |
| 2279 | * SCX ops can recurse via scx_bpf_sub_dispatch() - the inner call must not
|
| 2280 | * clobber the outer's scx_locked_rq_state. Save it on entry, restore on exit.
|
| 2281 | *
|
| 2282 | * @ops is the ops table to dispatch through: ops for the cpu form, ops_cid
|
| 2283 | * for the cid form.
|
| 2284 |
|
| 2285 | #define __SCX_CALL_OP(sch, ops, op, locked_rq, args...) \
|
| 2286 | do { \
|
| 2287 | struct rq *__prev_locked_rq; \
|
| 2288 | \
|
| 2289 | if (locked_rq) { \
|
| 2290 | __prev_locked_rq = scx_locked_rq(); \
|
| 2291 | update_locked_rq(locked_rq); \
|
| 2292 | } \
|
| 2293 | (sch)->ops.op(args); \
|
| 2294 | if (locked_rq) \
|
| 2295 | update_locked_rq(__prev_locked_rq); \
|
| 2296 | } while (0)
|
| 2297 |
|
| 2298 | #define SCX_CALL_OP(sch, op, locked_rq, args...) \
|
| 2299 | __SCX_CALL_OP(sch, ops, op, locked_rq, ##args)
|
| 2300 |
|
| 2301 | #define SCX_CALL_OP_RET(sch, op, locked_rq, args...) \
|
| 2302 | ({ \
|
| 2303 | struct rq *__prev_locked_rq; \
|
| 2304 | __typeof__((sch)->ops.op(args)) __ret; \
|
| 2305 | \
|
| 2306 | if (locked_rq) { \
|
| 2307 | __prev_locked_rq = scx_locked_rq(); \
|
| 2308 | update_locked_rq(locked_rq); \
|
| 2309 | } \
|
| 2310 | __ret = (sch)->ops.op(args); \
|
| 2311 | if (locked_rq) \
|
| 2312 | update_locked_rq(__prev_locked_rq); \
|
| 2313 | __ret; \
|
| 2314 | })
|
| 2315 |
|
| 2316 | |
| 2317 | * SCX_CALL_OP_TASK*() invokes an SCX op that takes one or two task arguments
|
| 2318 | * and records them in current->scx.kf_tasks[] for the duration of the call. A
|
| 2319 | * kfunc invoked from inside such an op can then use
|
| 2320 | * scx_kf_arg_task_ok() to verify that its task argument is one of
|
| 2321 | * those subject tasks.
|
| 2322 | *
|
| 2323 | * Every SCX_CALL_OP_TASK*() call site invokes its op with @p's rq lock held -
|
| 2324 | * either via the @locked_rq argument here, or (for ops.select_cpu()) via @p's
|
| 2325 | * pi_lock held by try_to_wake_up() with rq tracking via scx_rq.in_select_cpu.
|
| 2326 | * So if kf_tasks[] is set, @p's scheduler-protected fields are stable.
|
| 2327 | *
|
| 2328 | * kf_tasks[] can not stack, so task-based SCX ops must not nest. The
|
| 2329 | * WARN_ON_ONCE() in each macro catches a re-entry of any of the three variants
|
| 2330 | * while a previous one is still in progress.
|
| 2331 |
|
| 2332 | #define __SCX_CALL_OP_TASK(sch, ops, op, locked_rq, task, args...) \
|
| 2333 | do { \
|
| 2334 | WARN_ON_ONCE(current->scx.kf_tasks[0]); \
|
| 2335 | current->scx.kf_tasks[0] = task; \
|
| 2336 | __SCX_CALL_OP((sch), ops, op, locked_rq, task, ##args); \
|
| 2337 | current->scx.kf_tasks[0] = NULL; \
|
| 2338 | } while (0)
|
| 2339 |
|
| 2340 | |
| 2341 | * A per-task op runs on @task's owner - WARN if @sch isn't it. Sites that must
|
| 2342 | * target a different scheduler call __SCX_CALL_OP_TASK() directly.
|
| 2343 |
|
| 2344 | #define SCX_CALL_OP_TASK(sch, op, locked_rq, task, args...) \
|
| 2345 | do { \
|
| 2346 | WARN_ON_ONCE(scx_has_subs() && (sch) != scx_task_sched_rcu(task)); \
|
| 2347 | __SCX_CALL_OP_TASK((sch), ops, op, locked_rq, task, ##args); \
|
| 2348 | } while (0)
|
| 2349 |
|
| 2350 | |
| 2351 | * Dispatch a task op through the cid-form ops_cid table, for the ops whose
|
| 2352 | * cid-form signature differs from the cpu-form slot: set_cmask() takes an arena
|
| 2353 | * cmask instead of a cpumask and enable() takes scx_enable_args.
|
| 2354 |
|
| 2355 | #define SCX_CALL_CID_OP_TASK(sch, op, locked_rq, task, args...) \
|
| 2356 | __SCX_CALL_OP_TASK(sch, ops_cid, op, locked_rq, task, ##args)
|
| 2357 |
|
| 2358 | #define SCX_CALL_OP_TASK_RET(sch, op, locked_rq, task, args...) \
|
| 2359 | ({ \
|
| 2360 | __typeof__((sch)->ops.op(task, ##args)) __ret; \
|
| 2361 | WARN_ON_ONCE(scx_has_subs() && (sch) != scx_task_sched_rcu(task)); \
|
| 2362 | WARN_ON_ONCE(current->scx.kf_tasks[0]); \
|
| 2363 | current->scx.kf_tasks[0] = task; \
|
| 2364 | __ret = SCX_CALL_OP_RET((sch), op, locked_rq, task, ##args); \
|
| 2365 | current->scx.kf_tasks[0] = NULL; \
|
| 2366 | __ret; \
|
| 2367 | })
|
| 2368 |
|
| 2369 | #define SCX_CALL_OP_2TASKS_RET(sch, op, locked_rq, task0, task1, args...) \
|
| 2370 | ({ \
|
| 2371 | __typeof__((sch)->ops.op(task0, task1, ##args)) __ret; \
|
| 2372 | WARN_ON_ONCE(current->scx.kf_tasks[0]); \
|
| 2373 | current->scx.kf_tasks[0] = task0; \
|
| 2374 | current->scx.kf_tasks[1] = task1; \
|
| 2375 | __ret = SCX_CALL_OP_RET((sch), op, locked_rq, task0, task1, ##args); \
|
| 2376 | current->scx.kf_tasks[0] = NULL; \
|
| 2377 | current->scx.kf_tasks[1] = NULL; \
|
| 2378 | __ret; \
|
| 2379 | })
|
| 2380 |
|
| 2381 |
|
| 2382 | static __always_inline bool scx_kf_arg_task_ok(struct scx_sched *sch,
|
| 2383 | struct task_struct *p)
|
| 2384 | {
|
| 2385 | if (unlikely((p != current->scx.kf_tasks[0] &&
|
| 2386 | p != current->scx.kf_tasks[1]))) {
|
| 2387 | scx_error(sch, "called on a task not being operated on");
|
| 2388 | return false;
|
| 2389 | }
|
| 2390 |
|
| 2391 | return true;
|
| 2392 | }
|
| 2393 |
|
| 2394 | static inline bool scx_bypassing(struct scx_sched *sch, s32 cpu)
|
| 2395 | {
|
| 2396 | return unlikely(per_cpu_ptr(sch->pcpu, cpu)->flags &
|
| 2397 | SCX_SCHED_PCPU_BYPASSING);
|
| 2398 | }
|
| 2399 |
|
| 2400 | #ifdef CONFIG_EXT_SUB_SCHED
|
| 2401 | DECLARE_STATIC_KEY_FALSE(__scx_has_subs);
|
| 2402 |
|
| 2403 | |
| 2404 | * scx_has_subs - Whether any sub-scheduler exists
|
| 2405 | *
|
| 2406 | * Gates the sub-sched portions of hot paths so that a root-only system doesn't
|
| 2407 | * pay for them. See scx_sub_enable_workfn() and scx_sched_free_rcu_work().
|
| 2408 |
|
| 2409 | static inline bool scx_has_subs(void)
|
| 2410 | {
|
| 2411 | return static_branch_unlikely(&__scx_has_subs);
|
| 2412 | }
|
| 2413 |
|
| 2414 | |
| 2415 | * scx_task_sched - Find scx_sched scheduling a task
|
| 2416 | * @p: task of interest
|
| 2417 | *
|
| 2418 | * Return @p's scheduler instance. Must be called with @p's pi_lock or rq lock
|
| 2419 | * held.
|
| 2420 |
|
| 2421 | static inline struct scx_sched *scx_task_sched(const struct task_struct *p)
|
| 2422 | {
|
| 2423 | return rcu_dereference_protected(p->scx.sched,
|
| 2424 | lockdep_is_held(&p->pi_lock) ||
|
| 2425 | lockdep_is_held(__rq_lockp(task_rq(p))));
|
| 2426 | }
|
| 2427 |
|
| 2428 | |
| 2429 | * scx_task_sched_rcu - Find scx_sched scheduling a task
|
| 2430 | * @p: task of interest
|
| 2431 | *
|
| 2432 | * Return @p's scheduler instance. The returned scx_sched is RCU protected.
|
| 2433 |
|
| 2434 | static inline struct scx_sched *scx_task_sched_rcu(const struct task_struct *p)
|
| 2435 | {
|
| 2436 | return rcu_dereference_all(p->scx.sched);
|
| 2437 | }
|
| 2438 |
|
| 2439 | |
| 2440 | * scx_task_on_sched - Is a task on the specified sched?
|
| 2441 | * @sch: sched to test against
|
| 2442 | * @p: task of interest
|
| 2443 | *
|
| 2444 | * Returns %true if @p is on @sch, %false otherwise.
|
| 2445 |
|
| 2446 | static inline bool scx_task_on_sched(struct scx_sched *sch,
|
| 2447 | const struct task_struct *p)
|
| 2448 | {
|
| 2449 | return rcu_access_pointer(p->scx.sched) == sch;
|
| 2450 | }
|
| 2451 |
|
| 2452 | |
| 2453 | * scx_prog_sched - Find scx_sched associated with a BPF prog
|
| 2454 | * @aux: aux passed in from BPF to a kfunc
|
| 2455 | *
|
| 2456 | * To be called from kfuncs. Return the scheduler instance associated with the
|
| 2457 | * BPF program given the implicit kfunc argument aux. The returned scx_sched is
|
| 2458 | * RCU protected.
|
| 2459 |
|
| 2460 | static inline struct scx_sched *scx_prog_sched(const struct bpf_prog_aux *aux)
|
| 2461 | {
|
| 2462 | struct sched_ext_ops *ops;
|
| 2463 | struct scx_sched *sch, *root;
|
| 2464 |
|
| 2465 | ops = bpf_prog_get_assoc_struct_ops(aux);
|
| 2466 | if (likely(ops)) {
|
| 2467 | sch = rcu_dereference_all(ops->priv);
|
| 2468 | if (sch && unlikely(READ_ONCE(sch->dead)))
|
| 2469 | return NULL;
|
| 2470 | return sch;
|
| 2471 | }
|
| 2472 |
|
| 2473 | root = rcu_dereference_all(scx_root);
|
| 2474 | if (root) {
|
| 2475 | if (unlikely(READ_ONCE(root->dead)))
|
| 2476 | return NULL;
|
| 2477 | |
| 2478 | * COMPAT-v6.19: Schedulers built before sub-sched support was
|
| 2479 | * introduced may have unassociated non-struct_ops programs.
|
| 2480 |
|
| 2481 | if (!root->ops.sub_attach)
|
| 2482 | return root;
|
| 2483 |
|
| 2484 | if (!root->warned_unassoc_progs) {
|
| 2485 | printk_deferred(KERN_WARNING "sched_ext: Unassociated program %s (id %d)\n",
|
| 2486 | aux->name, aux->id);
|
| 2487 | root->warned_unassoc_progs = true;
|
| 2488 | }
|
| 2489 | }
|
| 2490 |
|
| 2491 | return NULL;
|
| 2492 | }
|
| 2493 |
|
| 2494 | |
| 2495 | * scx_parent - Find the parent sched
|
| 2496 | * @sch: sched to find the parent of
|
| 2497 | *
|
| 2498 | * Returns the parent scheduler or %NULL if @sch is root.
|
| 2499 |
|
| 2500 | static inline struct scx_sched *scx_parent(struct scx_sched *sch)
|
| 2501 | {
|
| 2502 | if (sch->level)
|
| 2503 | return sch->ancestors[sch->level - 1];
|
| 2504 | else
|
| 2505 | return NULL;
|
| 2506 | }
|
| 2507 |
|
| 2508 | #else
|
| 2509 | static inline bool scx_has_subs(void) { return false; }
|
| 2510 |
|
| 2511 | static inline struct scx_sched *scx_task_sched(const struct task_struct *p)
|
| 2512 | {
|
| 2513 | return rcu_dereference_protected(scx_root,
|
| 2514 | lockdep_is_held(&p->pi_lock) ||
|
| 2515 | lockdep_is_held(__rq_lockp(task_rq(p))));
|
| 2516 | }
|
| 2517 |
|
| 2518 | static inline struct scx_sched *scx_task_sched_rcu(const struct task_struct *p)
|
| 2519 | {
|
| 2520 | return rcu_dereference_all(scx_root);
|
| 2521 | }
|
| 2522 |
|
| 2523 | static inline bool scx_task_on_sched(struct scx_sched *sch,
|
| 2524 | const struct task_struct *p)
|
| 2525 | {
|
| 2526 | return true;
|
| 2527 | }
|
| 2528 |
|
| 2529 | static inline struct scx_sched *scx_prog_sched(const struct bpf_prog_aux *aux)
|
| 2530 | {
|
| 2531 | struct scx_sched *root = rcu_dereference_all(scx_root);
|
| 2532 |
|
| 2533 | if (root && unlikely(READ_ONCE(root->dead)))
|
| 2534 | return NULL;
|
| 2535 | return root;
|
| 2536 | }
|
| 2537 |
|
| 2538 | static inline struct scx_sched *scx_parent(struct scx_sched *sch) { return NULL; }
|
| 2539 |
|
| 2540 | #endif
|
| 2541 |
|
| 2542 | #endif
|